GRALIS-Report: auditable region-level attribution and structured clinical report generation for breast cancer histology

1 Introduction

1.1 The clinical problem

Histological diagnosis of breast cancer requires years of specialized training. Deep learning models now achieve accuracy comparable to that of expert pathologists on digitized biopsies (Spanhol et al., 2016; Litjens et al., 2017), yet clinical adoption remains limited by a fundamental obstacle: when a model labels an image as malignant, it does not indicate which tissue structures it used to reach that conclusion. A pathologist cannot validate—or challenge—a decision they cannot inspect (Rudin, 2019).

1.2 What GRALIS does

GRALIS (Gradient-Riesz Averaged Locally-Integrated Shapley) is an Explainable AI method that answers this question directly. Given a histological image and a classifier's prediction, GRALIS partitions the image into tissue segments (superpixels) and assigns each segment an importance score ϕ_*i* ∈ ℝ: a positive score indicates that the segment increased the malignancy probability; a negative score means it decreased it. The result is a semantic attribution signal—a piecewise-constant map over tissue segments, visualizable as a heat map, that encodes the spatial distribution of model evidence in the region-level format intended to support inspection by pathologists.

1.3 How GRALIS improves on existing methods

GRALIS is designed to occupy a specific point in the XAI design space: superpixel-level spatial coherence [addressing GradCAM's resolution limitations and IG's perceptually uninformative dense pixel maps (Arun et al., 2021)], an explicit a priori Monte Carlo sample-size bound (Fanale, 2026) [addressing KernelSHAP/LIME's reliance on post-hoc convergence detection (Covert and Lee, 2021)], and a non-zero empirical completeness residual as a deliberate trade-off for LIME-style locality. No single method satisfies all desiderata; the contribution of GRALIS is that it jointly satisfies the three properties above—a combination not available in existing methods—at the cost of marginal pixel-level faithfulness compared with IG across independent metrics. Table 1 summarizes which properties each method satisfies (detailed comparison in Section 2). Table 2 defines the intended scope and non-claims of this study.

Table 1. Theoretical property comparison.

✓ = property holds by design; ≈ = approximately satisfied as a deliberate trade-off; × = property absent or requires additional assumptions. Complete: attribution scores exactly sum to the model output change (∑iϕi=f(x)-f(x′)). A priori bound: the required number of perturbation samples can be determined before running, given a target approximation error ε [as opposed to post-hoc convergence detection (Covert and Lee, 2021)]. All methods converge asymptotically; the column distinguishes whether a sample-size bound is available a priori. For GRALIS, ≈ indicates that exact completeness does not hold for the reported unnormalized attribution values; the empirical residual is quantified separately (Section 4.7). The non-uniform LIME kernel breaks the Shapley telescoping sum as a deliberate trade-off for attribution locality; completeness for the normalized game is established in the companion theoretical work (Fanale, 2026).

^*KernelSHAP is a consistent estimator of Shapley values; × denotes absence of an a priori sample-size bound, not lack of convergence (Covert and Lee, 2021).

Table 2. Intended scope of this paper.

Left column: what this paper is and claims. Right column: what it does not claim.

1.4 Main contributions

GRALIS builds on a mathematical framework developed in a companion paper (Fanale, 2026) (arXiv:2605.05480); proofs and theoretical derivations are fully contained therein. The present study is entirely experimental: it implements GRALIS on a GPU cluster, validates it on 1,187 breast histology images, benchmarks it against competing methods, and integrates the output into a structured, pathology-oriented reporting prototype.

This study makes the following contributions:

• GRALIS-Report: an auditable cross-modal attribution pipeline. Raw images enter the attribution stage but not the report generation stage; the only input to reporting is the structured attribution summary F(x) (superpixel scores, confidence, classification outcome). This explicit symbolic intermediate layer enforces a clean separation between visual evidence and pathology-oriented language, enabling auditability, offline operation, and deterministic reproducibility.
• GRALIS: coalition-conditioned attribution with a priori convergence control. The novel algorithmic element is coalition-conditioned path integration: the gradient is evaluated along a path specific to each coalition S, capturing model curvature not present, to our knowledge, in LIME, KernelSHAP, or standard IG (formal definition in Appendix). A direct consequence is a pre-run sample-size bound computable from the LIME kernel and gradient norms across the benchmarked methods. None provides the same pre-run Monte Carlo bound for the locality-weighted, coalition-conditioned formulation used here [proof in Fanale (2026)].
• A faithfulness benchmark on the full BreaKHis test set (1,187 images): pixel-level deletion AUC (Petsiuk et al., 2018) for all six methods; ROAD (Rong et al., 2022) MoRF/LeRF evaluation for three methods (GRALIS, IG, GradCAM); and a granularity-fair superpixel masking protocol as primary metrics. ExpiScore_*s* is reported as a supplementary profile metric; results on it should be read with the declared conflict of interest in mind.
• An expert perception discordance study (N = 4 anatomopathologists, 60 cases): marked inter-rater variability on perceived clinical utility (individual Q13 means: 1.47–4.73/5) suggests that perceived explanation utility may not constitute a stable ground truth in this domain. This result is reported as a methodological finding with implications for the design of future XAI human-evaluation protocols in histopathology, not as evidence for or against clinical utility.

2 Related work

2.1 XAI for medical imaging

GradCAM (Selvaraju et al., 2017) is widely used in medical imaging for its computational simplicity, but it produces low-resolution maps (7 × 7 for DenseNet) that require bilinear interpolation. Furthermore, the gradient averaging step can highlight regions the model did not actually use (Draelos and Carin, 2020). KernelSHAP (Lundberg and Lee, 2017) and LIME (Ribeiro et al., 2016) operate on superpixels but lack a priori sample-size bounds; KernelSHAP relies on post-hoc convergence detection to determine when enough samples have been drawn (Covert and Lee, 2021), yielding variable approximation quality across images. Integrated Gradients (Sundararajan et al., 2017) guarantees completeness in pixel space but produces dense per-pixel maps (SAL = 0.150 in our benchmark) that may be difficult to relate to clinically meaningful tissue structures in medical imaging (Arun et al., 2021). A systematic review of XAI methods in digital oncology is provided in Arrieta et al. (2020); Reyes et al. (2020). GRALIS addresses these limitations jointly: superpixel-space operation, an explicit Monte Carlo sample-size bound [Theorem T3 in Fanale (2026)], and a non-zero empirical completeness residual (|∑iϕi-Δf|≈0.40, a deliberate design consequence of the LIME locality kernel). The completeness trade-off is deliberate: a non-constant locality kernel is formally incompatible with exact efficiency [Proposition 1 in Fanale (2026)], and preserving locality is the more important property for region-level inspectability.

2.2 Why superpixels matter clinically

Pathological assessment of histological tissue is inherently region-based. Pathologists identify candidate regions—cell clusters, ductal structures, stromal areas—and reason about their morphological properties, not about individual pixels (Gurcan et al., 2009). A pixel-level attribution map (as produced by Integrated Gradients or raw gradient methods) assigns importance to individual pixels, producing dense maps whose spatial extent does not correspond to any anatomically meaningful structure and which trained experts find difficult to relate to tissue architecture (Arun et al., 2021).

Superpixel-level attribution addresses this directly: each attribution unit ϕ_*i* corresponds to a spatially coherent region that a pathologist can inspect, describe, and discuss. GRALIS-Report is designed around this human-centered framing: the primary criterion for map quality is alignment with the pathologist's region-based reasoning (region-level inspectability), with faithfulness to the model reported as a complementary, independently measured dimension (Section 3.5).

2.3 Knowledge distillation for histological classification

The DenseNet architecture (Huang et al., 2017) has demonstrated excellent performance on BreaKHis (Spanhol et al., 2016). Knowledge distillation (Hinton et al., 2015) transfers the “knowledge” of a larger teacher model (InceptionV3) into a more compact student (DenseNet-121), maintaining accuracy with reduced computational complexity.

2.4 Automated clinical report generation

Automated generation of radiological and pathological reports has emerged as a research frontier (Thirunavukarasu et al., 2023; Moor et al., 2023). Integrating XAI maps into report generation for visual grounding of explanations has been explored in radiology (Bannur et al., 2023), but not in breast histology with an explicit attribution intermediate layer. Our report generation module uses a deterministic, rule-based approach that ensures reproducibility and transparency without relying on external language model APIs.

2.5 Relationship to the ExpiScore evaluation framework

ExpiScore (Fanale et al., 2026) and the present study are methodologically distinct: ExpiScore defines how to measure explanation quality across existing methods; GRALIS defines what to compute as a new attribution method. A conflict of interest exists: the author of this study is also the first author of Fanale et al. (2026), and ExpiScore rewards properties that superpixel-based maps inherently possess. To address this, deletion AUC and ROAD—both independent of the author—are the primary faithfulness metrics here; ExpiScore_*s* is supplementary.

This study uses ExpiScore_*s* = (SAL + CPT + CSC)/3, a simplified variant that omits the training-set normalization and OGSL consensus of the full formula in Fanale et al. (2026); values are not numerically comparable across the two papers.

Table 1 summarizes GRALIS against the four benchmark methods with respect to the theoretical properties most relevant to clinical deployment.

A method is “complete” if its attribution scores sum exactly to the model's output change; “convergent” if its approximation error decreases at a known rate; “superpixel-based” if it produces segment-level maps suitable for histological interpretation; “resolution-independent” if it does not depend on the size of convolutional feature maps.

3 Materials and methods

3.1 Dataset and preprocessing

We use the BreaKHis dataset (Spanhol et al., 2016), comprising 7,909 histological images of breast tissue acquired at four magnifications (40×, 100×, 200×, 400×) with H&E staining, at a resolution of 700 × 460 pixels. Images span 8 histological subtypes: 4 benign (adenosis, fibroadenoma, phyllodes tumor, and tubular adenoma) and 4 malignant (ductal carcinoma, lobular carcinoma, mucinous carcinoma, and papillary carcinoma).

The adopted split yields: 5,536 training images, 1,186 validation images, and 1,187 test images (372 benign, 815 malignant in the test set). The split is performed at the patient level: all images from a given patient appear exclusively in one partition, preventing patient-overlap leakage—a known source of inflated performance in BreaKHis studies (Spanhol et al., 2016). Stratification is applied jointly by histological subtype and magnification factor to ensure balanced class representation across splits. Preprocessing follows the DenseNet protocol: resizing to 224 × 224, normalization by ImageNet mean/standard deviation. Test images are used exclusively for final evaluation (no data augmentation, no hyperparameter tuning on test set).

3.2 Classifier model and cross-dataset adaptation

3.2.1 Main classification model (BreaKHis, in-domain)

The classifier is a DenseNet-121 student (7, 304, 257 parameters) trained via knowledge distillation from an InceptionV3 teacher (22, 925, 857 parameters), with a binary sigmoid output (benign/malignant). Training uses Adam (η = 10⁻⁴, β₁ = 0.9, β₂ = 0.999), A batch size of 32, early stopping on validation AUC (patience = 10, temperature T = 2.0, distillation weight α = 0.5). The BreaKHis dataset was split into train (5,536 images, 70%), validation (1,186 images, 15%), and held-out test (1,187 images, 15%) sets, with no image-level overlap between splits. On the held-out BreaKHis test set, the student achieved AUC = 0.9989, accuracy = 0.9916, sensitivity = 0.9939, and specificity = 0.9866, surpassing the InceptionV3 teacher (AUC = 0.9962, accuracy = 0.9714). The trained DenseNet-121 backbone (conv5_block16_2_conv as the last convolutional layer) serves as the feature extractor for all XAI methods in both the in-domain and cross-dataset evaluations. Performance is also reported in Table 3.

Table 3. Performance of InceptionV3 teacher and DenseNet-121 student on the BreaKHis test set (knowledge distillation, α = 0.5, T = 2.0).

Figures reflect a patient-level split on a single-institution dataset and represent internal benchmark performance only; they should not be interpreted as evidence of generalizability or deployment readiness.

3.2.2 Cross-dataset linear probe (IDC)

To evaluate generalization to a different imaging protocol, we adopted a transfer learning protocol with a frozen backbone: the DenseNet-121 student weights trained on BreaKHis were frozen entirely, and only a two-layer classification head (Dropout(0.3) + Dense(1, sigmoid), ≈1, 025 trainable parameters, < 0.02% of total model parameters) was trained on the IDC Breast Cancer dataset (Janowczyk and Madabhushi, 2016; Cruz-Roa et al., 2014) (Penn Medicine/Kaggle, CC0 license). The IDC training set consisted of 400 benign and 400 malignant 50 × 50px patches; the test set comprised 249 benign and 250 malignant patches (499 total). Class weights were computed from training class frequencies. The probe was optimized with Adam (lr = 10⁻³, batch size 32) with early stopping on validation AUC (patience = 10) and learning rate reduction on plateau (factor 0.5, patience = 5); training converged at epoch 29. The IDC probe was re-trained under a strict held-out protocol (internal 80/20 train/val split; 499-image test set never accessed during training or epoch selection). Test AUC changed by only Δ = −0.0015 (0.8977 → 0.8962); all attribution maps were regenerated; the GRALIS ranking is identical (rank 1/6 on both metrics). On the IDC test set, the clean-run linear probe achieved AUC = 0.8962 and accuracy = 0.8116. This constitutes a cross-dataset transfer setting with a frozen BreaKHis-trained feature extractor and a target-domain linear probe; the attribution mechanism operates on backbone features that were never trained on IDC data.

3.2.3 Cross-dataset linear probe (PCam)

The PatchCamelyon dataset (Veeling et al., 2018) (HuggingFace 1aurent/PatchCamelyon, CC0 license) provides 96 × 96 px H&E-stained lymph node patches with binary labels (tumor present in The central 32 × 32 region). The training set comprised 400 benign and 400 malignant patches (800 total), randomly sampled from the official PCam training split with seed = 42; the evaluation set comprised 250 benign and 250 malignant patches (500 total) randomly sampled from the official PCam test split with seed = 42. All images were resized to 224 × 224 px (LANCZOS) and preprocessed with tf.keras.applications.densenet.preprocess_ input, identical to the BreaKHis/IDC pipeline. No data augmentation was applied. The probe architecture, training configuration, and hyperparameters are identical to the IDC protocol: Dropout(0.3) + Dense (1, sigmoid) (≈1, 025 trainable parameters, < 0.02% of total model parameters), Adam (lr = 10⁻³, batch size 32), class weights computed from training frequencies (balanced dataset: weights effectively equal), early stopping on validation AUC (patience = 10, restore best weights), and learning rate reduction on plateau (factor 0.5, patience = 5, minimum lr = 10⁻⁶); maximum 50 epochs. Training converged at epoch 30. Validation protocol: an internal 80/20 train/validation split (640 train, 160 val) was applied to the 800 official PCam training images; the 500 evaluation images (official PCam test split) were held out entirely during training and epoch selection. Probe hyperparameters were fixed a priori based on the IDC configuration; no PCam-specific hyperparameter search was performed. On the held-out PCam evaluation set, the linear probe achieved AUC = 0.8815 and the same frozen backbone and probe were used for all six XAI attribution methods without modification.

3.3 GRALIS Algorithm 1—implementation

GRALIS Algorithm 1 (Fanale, 2026) is implemented in TensorFlow 2 with GPU optimization for A100-80GB (pipeline overview: Figure 1). Adopted parameters: n_seg = 30 SLIC superpixels (Achanta et al., 2012) (compactness = 50, σ = 1.0; effective: 25 ± 2 per image), m = 100 Monte Carlo permutations with antithetic variants (200 effective samples), k = 50 integration steps (midpoint rule, error O(1/k²)), σ_LIME = 0.75, baseline x′ = 0.

3.3.1 GPU optimizations

The full batch of size n_seg × k per permutation (up to 250 images) is built using two vectorized NumPy operations (broadcast over α and cumulative coalition masks), eliminating Python loops over images. A single GradientTape per permutation processes the entire batch. Antithetic pairs (τ, τ⁻¹) are included in the same doubled batch to reduce variance without computational overhead.

3.3.2 Piecewise-constant attribution map

Unlike standard IG, which produces continuous pixel-level attributions, GRALIS Algorithm 1 computes for each position j in the permutation. Notation: τ is a permutation of the n_seg superpixels; S_*j* = {τ(1), …, τ(j − 1)} is the coalition of segments appearing before position j; x~(Sj,α) is the α-interpolated image with pixels in S_*j* set to xhw′+α(xhw-xhw′) and all other pixels held at the zero baseline x′; f denotes the classifier output (scalar in [0, 1]); k is the number of integration steps (midpoint rule):

IGτ(j)Sj=∑(h,w)∈segτ(j)(xhw-xhw′)·1k∑α∂f∂xhw(x~(Sj,α)),(1)

and assigns this scalar as the uniform attribution to all pixels in segment τ(j). The resulting map has at most n_seg distinct values, ensuring by construction the sparsity required for high compactness.

3.3.3 Sample-size bound

GRALIS provides a theoretical a priori Monte Carlo sample-size bound [Theorem T3 in Fanale (2026)], whose required constant B is computable from the LIME kernel and gradient norms for a target approximation error ε. Numerical computation of B requires a calibration pass through the attribution pipeline to estimate gradient norms; it can be computed before the full Monte Carlo attribution run, after a calibration pass. The present experiments do not constitute numerical validation of the bound because the dataset-specific constant B was not estimated. The reported 29% variance reduction concerns antithetic sampling only and should not be interpreted as empirical verification of the sample-size guarantee.

3.3.4 Computational environment

BreaKHis experiments were executed on a single NVIDIA A100-80GB GPU (CUDA 12.x, TensorFlow 2). Cross-dataset experiments (IDC and PCam) were executed on a single NVIDIA H100 SXM5 80 GB GPU (CUDA 12.8, cuDNN 9.1) with 64 CPU cores and 512 GB RAM. For IDC: attribution map generation for the five baselines over 499 test images required approximately 6.5 h (≈47 s per image; KernelSHAP and LIME dominating with 500 masks each); GRALIS-MC (m = 100, k = 50) required approximately 5.5 h (≈40 s per image); faithfulness evaluation (499 × 6 = 2, 994 evaluations) completed in under 1 h; total ≈13 h. For PCam: GRALIS-MC attribution (m = 100, k = 50) required approximately 20 min (≈2.4 s per image; all images resized to 224 × 224 before inference, identical to IDC); faithfulness evaluation completed in approximately 30 min; total ≈1.5 h including probe training. Note on PCam/IDC timing: both figures were obtained in separate runs and should not be interpreted as a controlled cross-dataset timing comparison.

Method Summary (self-contained statements).

(1) Attribution definition. For superpixel i, GRALIS computes:

ϕi=1m∑τωx(Sτ(i))·IGτ(i)Sτ(i)

where the sum is over m Monte Carlo permutations τ, S_τ(*i*) is the coalition of segments preceding i in τ, ωx(S)=exp(-‖xS-xS′‖2/2σ2) is the LIME locality kernel, and IGτ(i)Sτ(i) is the coalition-conditioned integrated gradient for segment τ(i) (Equation 1).

(2) Completeness gap. Exact completeness is not expected under the non-uniform LIME kernel [Theorem T2, Fanale (2026)]; empirically, the observed gap is |∑iϕi-(f(x)-f(x′))|≈0.40±0.18. This is a deliberate design trade-off: the locality kernel ensures region-coherent attributions at the cost of exact Shapley efficiency.

(3) Convergence bound. Theorem T3 (Fanale, 2026) establishes |ϕ^i(m,k)-ϕiGRALIS|≤B/mδ+‖∇2f‖∞(xi-xi′)2/(8k), where B=supS,α|ωx(xS)(xi-xi′)∇if(x~(S,α))| is computable before the full Monte Carlo attribution run, after a calibration pass. Antithetic sampling halves the MC variance (proved in T3): a variance factor of 1/2 corresponds to a 1-1/2≈29.3% reduction in standard deviation, consistent with the empirically observed 29% σ reduction (Section 4.7).

3.4 XAI baseline methods

We compare GRALIS with five baseline methods, all applied to the same DenseNet-121 model:

• GradCAM (Selvaraju et al., 2017): gradients from the final convolutional layer, bilinear upsampling to 224 × 224.
• GradCAM++ (Chattopadhay et al., 2018): higher-order gradient weighting applied to the final convolutional layer, followed by bilinear upsampling to 224 × 224; generalizes GradCAM by using squared gradients and a normalization factor per channel.
• KernelSHAP (Lundberg and Lee, 2017): 500 samples, SLIC segmentation with n = 30 superpixels, Shapley kernel.
• LIME (Ribeiro et al., 2016): 1,000 perturbations, SLIC segmentation with n = 30 superpixels, local linear regressor.
• Integrated gradients (Sundararajan et al., 2017): 50 steps, baseline x′ = 0, mean of gradients along the linear path.

3.5 Evaluation metrics

3.5.1 Three complementary evaluation dimensions

This study evaluates attribution quality along three conceptually distinct dimensions that should not be conflated:

• Perceptual map properties/region-level inspectability: whether the map is spatially concentrated on salient regions rather than diffuse noise, and whether it is compact enough to support region-based cognitive assessment by a pathologist. Captured primarily by SAL (salience concentration) and CPT (compactness), which are proxies for concentration and compactness, not direct measures of clinical interpretability; and indirectly by the human preference study in Section 4.9.
• Faithfulness to the model: whether attributions accurately reflect which inputs actually influenced the model's prediction. Captured by deletion AUC (Petsiuk et al., 2018) and ROAD MoRF (Rong et al., 2022).
• Cross-method agreement: whether independent attribution methods assign similar scores to the same regions. Captured by CSC (cross-method consistency score, Section 2.5). Agreement is a necessary but not sufficient condition for faithfulness: methods may collectively agree and be collectively wrong, or a faithful method may diverge from less faithful ones.

A map can score high on perceptual properties without being faithful (e.g., LIME's SAL≈1 reflects map saturation rather than genuine concentration). A map can also show high cross-method agreement without being perceptually inspectable—for example, IG has the highest CSC (0.562) but low SAL (0.150), rendering its maps perceptually uninformative for region-level inspection. Faithfulness is evaluated separately using Deletion AUC and ROAD MoRF. GRALIS is designed to balance perceptual and faithfulness axes, accepting A moderate faithfulness loss relative to pixel-precise methods in exchange for the spatial coherence that supports pathologist inspection.

3.5.2 ExpiScore

We use the ExpiScore framework (Fanale et al., 2026) with three components. Transparency note: ExpiScore was proposed by Fanale et al. (2026), a paper co-authored by the author of this work (now published as Frontiers in Imaging 5, 1846414). Because the metric rewards properties inherent to superpixel maps, results should be read alongside the independent deletion faithfulness test in Section 4.4 and the pixel-level deletion AUC (Petsiuk et al., 2018) in Table 4. Deletion AUC (Petsiuk et al., 2018) is reported for all six methods on the full test set in Tables 4, 5.

SAL=1|M|∑(h,w)∈Mϕ(h,w),M={(h,w):ϕ(h,w)≥P80(ϕ)}(2)

CPT=1-H(ϕ)Hmax, H(ϕ)=-∑pϕp‖ϕ‖1logϕp‖ϕ‖1(3)

CSC=2M(M-1)∑i<jDice(Mi,Mj),(4)

where M is the number of methods and M_*i* is the top-20% mask of method i. Throughout this study we use the notation

ExpiScores=SAL+CPT+CSC3,

where the subscript s denotes the simplified variant adopted here to distinguish it from the full normalized formula in Fanale et al. (2026). Formula note: The full ExpiScore in Fanale et al. (2026) uses min-max normalization, training-set statistics, and an OGSL consensus score; ExpiScore_*s* omits these, so numerical values are not directly comparable across the two papers.

Table 4. ExpiScore comparison on BreaKHis test set (1,187 images, complete).

Bold = best per column. ^*LIME SAL near 1 reflects a known property of binary perturbation schemes (Ribeiro et al., 2016): all superpixels receive binary importance values, causing the top-20% mask to saturate regardless of input content. ^†Faithfulness = Pearson correlation (feature removal vs. prediction drop). ^‡Empirical completeness residual (zero baseline) =|∑iϕi-Δf|, where Δf = f(x) − f(0); lower is better. For IG this is 0 by the completeness axiom (theoretical guarantee, not empirical); for other methods computed only where attribution maps are available in standardized input space (N/A otherwise). ^§Deletion AUC (Petsiuk et al., 2018) computed via pixel-level deletion on the full test set (N = 1, 187) for all six methods (lower = more faithful). Previous version reported a 100-image subsample for GRALIS, IG, and GradCAM only; full-set values are consistent with those estimates within one standard error.

Table 5. Deletion AUC for all six XAI methods on the complete BreaKHis test set (N = 1, 187).

Pixels are removed in descending attribution order at fractional steps [0, 0.05, 0.10, …, 1.0] and replaced with a zero baseline (ImageNet-normalized); AUC is computed with the trapezoidal rule. Lower AUC = faster confidence erosion = better faithfulness (Petsiuk et al., 2018). Wilcoxon signed-rank p-values (two-sided) compare each method against GRALIS; all differences are significant at p < 10⁻¹². Bold indicates the best (lowest) value per column.

3.5.3 Superpixel deletion faithfulness

We define the superpixel deletion drop as:

FDK=1N∑i=1N[f(xi)-f(xi\K)],(5)

where xi\K is the image with the top-K superpixels by mean attribution zeroed out. Unlike pixel-level deletion, this test is appropriate for piecewise-constant maps: the perturbation removes entire superpixels (≈2, 000 pixels each), producing semantically meaningful perturbations for the model.

3.6 GRALIS-report: two-tier clinical report architecture

GRALIS-Report implements a sequential cross-modal signal transformation:

histological image︸visual signal →classifier+ GRALIS ℱ(x)︸attribution signal →report engine SOAP report︸structured language

The raw image and its perturbed variants remain confined to the visual attribution stage. Once the structured attribution summary F(x) is produced, the report-generation stage operates exclusively on F(x)—a structured numerical representation of the classifier's evidence. This design decouples the visual processing stage from the language generation stage, enabling full auditability of each modality transition.

GRALIS-Report is built on three shared upstream components:

• Classifier: DenseNet-121 student producing binary prediction ŷ ∈ {0, 1} and confidence f(x) ∈ [0, 1].
• Attribution maps: GRALIS Algorithm 1 producing raw signed scores ϕiGRALIS∈ℝ per superpixel, normalized per image via min-max scaling to ϕnorm∈[0,1]224×224 piecewise constant for visualization and report generation (raw scores preserved in .npy files), plus segmentation map segs∈{0,…,ns}224×224.
• Structured findings extractor: Derives, from the run CSV and .npy files, the following attribution summary F(x): predicted label + confidence f(x); classification outcome (TP/TN/FP/FN, available in evaluation mode only, when ground-truth labels are provided; in prospective mode, only the predicted label and confidence are available); effective superpixel count n_*s*; mean attribution ϕ̄ and completeness proxy; top-K superpixels with area percentage and ϕ value; discordance flag.

F(x) is the sole input to both reporting modes, ensuring that the raw image never enters the report generation pipeline and that all structured model outputs are grounded in auditable attribution statistics.

3.6.1 Mode A—deterministic SOAP engine (primary)

A rule-based module translates F(x) into a SOAP-structured report (Subjective/Objective/Assessment/Plan) using a fixed deterministic vocabulary: confidence levels (high f ≥ 0.85, moderate 0.60 ≤ f < 0.85, low f < 0.60); attribution quality derived from ϕ̄ and completeness (concentrated/moderate/diffuse); a low-confidence advisory flag for f < 0.60 (applicable in both modes); in evaluation mode (ground-truth label available), an additional priority review alert for FP/FN cases. Identical input → identical output (full reproducibility). Mode A is the primary mode for all benchmarking, scientific claims, and audit traceability in this study.

3.6.2 Mode B—LLM narrative renderer (optional, not evaluated)

An optional module passes F(x)—not the raw image—to a large language model with the Mode A SOAP text as a structured scaffold. The LLM is constrained to enrich narrative style only and may not modify the structured model outputs encoded in F(x). Mode B is an architectural design concept; no quantitative evaluation is reported in this study, and it contributes to no result.

3.6.3 Clinical role

GRALIS-Report is designed to support secondary review and interpretability analysis; it is not intended to replace the diagnosis issued by a qualified anatomopathologist. All SOAP outputs should be treated as decision-support artifacts and are subject to pathologist review and clinical judgment at all times. Table 6 summarizes the content of each SOAP field.

Table 6. Structure of SOAP fields generated by GRALIS-Report.

The table describes evaluation-mode outputs. In prospective mode, ground truth, TP/TN/FP/FN status, and discordance-based priority alerts are unavailable.

4 Results

4.1 Classifier performance

Table 3 summarizes the performance of teacher and student models on the BreaKHis test set. The DenseNet-121 student exceeds the InceptionV3 teacher on all metrics after knowledge distillation, achieving 99.2% accuracy and AUC = 0.9989. The student exceeding the teacher (99.2% vs. 97.1% accuracy; AUC 0.9989 vs. 0.9962) is consistent with findings in knowledge distillation for medical imaging (Han et al., 2017; Huang et al., 2017). The observed advantage may reflect regularization from distillation, architecture-specific inductive biases, or run-to-run variation; the single-run design does not permit these factors to be separated. Both figures are internal benchmarks on the same test set; external validation on independent cohorts is required before clinical performance claims can be made. The comparison reflects a single training run; no ablation without knowledge distillation was performed, so the contribution of KD to the performance advantage over the teacher cannot be isolated from architectural effects. This comparison is not evidence that the DenseNet-121 student systematically outperforms InceptionV3 across runs or datasets.

4.2 Properties of GRALIS attribution maps

Table 7 reports statistics of attribution maps produced by GRALIS Algorithm 1 on all 1,187 test images. Qualitative examples are shown in Figure 2.

Table 7. GRALIS attribution map properties on BreaKHis. (1,187 images, complete test set).

The low CPT value (0.018) follows directly from the piecewise-constant map structure: with only 25 distinct values spread across 50,176 pixels, normalized pixel-level entropy is necessarily high regardless of semantic sparsity. The more appropriate sparsity proxy is ϕ_active = 0.35: only 35% of pixels belong to superpixels with attribution above the median, indicating concentration under this sparsity proxy.

4.3 ExpiScore comparison with baseline methods

Table 4 reports ExpiScore results for all methods on the full 1,187-image test set. Figure 3 visualizes the three components as a radar chart; Figure 4 shows the per-image ExpiScore distribution against classifier confidence.

Table 4 shows that no single method dominates all criteria. LIME achieves the highest overall ExpiScore due to SAL ≈ 1, but this reflects map saturation from its binary surrogate scheme rather than genuine selectivity (Ribeiro et al., 2016). IG achieves the highest cross-method consistency (CSC = 0.562), but SAL = 0.150 makes its maps perceptually uninformative. GRALIS shows SAL = 0.762 (third among the six compared methods), CPT = 0.018 (second), CSC = 0.274 (fourth), and deletion AUC = 0.704 (second out of six on the full test set, Table 5)—a profile that avoids the pronounced single-criterion weaknesses observed in the other methods within this comparison (ExpiScore conflict of interest declared in Section 2.5). GRALIS exhibits a mean empirical completeness residual of 0.40 ± 0.18 (higher than IG's theoretical zero) in exchange for spatial coherence, associated with the locality-weighted superpixel attribution formulation (Fanale, 2026).

4.4 Faithfulness: superpixel deletion

Table 8 reports superpixel deletion faithfulness results (n = 25 per class). The pattern is symmetric and theoretically consistent. For malignant images: removing the top-3 superpixels reduces malignant confidence in 96% of cases (drop = +0.025, σ = 0.058). For benign images: the drop is consistently negative (96% of cases), indicating that GRALIS identifies superpixels that suppress the malignant signal; their removal pushes the model toward malignant, supporting sign-consistent model faithfulness under this perturbation protocol.

Table 8. Superpixel deletion faithfulness (n = 25/class). Drop = f(x)−f(x_\*K*).

Bold indicates the best (lowest) value per column.

This result motivates the definition of a class-conditional faithfulness metric: FDK+ for the target class and FDK- for the complementary class, separating positive and negative contributions to prediction.

4.5 Multi-method faithfulness benchmark

To contextualize GRALIS faithfulness among established XAI methods, we benchmark all six methods on the complete BreaKHis test set (N = 1, 187) using two complementary protocols applied to pre-computed attribution maps.

Pearson faithfulness measures the Pearson correlation between the fraction of pixels removed (ranked by descending attribution) and the cumulative drop in model confidence (Samek et al., 2016). Higher values indicate that removing the most attributed pixels produces the largest and most monotone confidence reduction.

Pixel-level masking faithfulness replaces the top-p% of pixels (by attribution magnitude) with a neutral gray fill (128/255) and records Δp=f(xmasked)-f(x). A more negative Δ_*p* indicates faster model-confidence erosion and hence stronger pixel-level faithfulness (Petsiuk et al., 2018).

Table 9 reports both metrics for all six methods.

Table 9. Multi-method faithfulness benchmark on the complete BreaKHis test set (N = 1, 187).

Pearson: Pearson correlation between removal fraction and confidence drop (higher = better); rank is by Pearson. Δ@p%: mean ± std confidence change when the top-p% of pixels by attribution magnitude are replaced with neutral gray (more negative = better). Wilcoxon signed-rank p-values compare each method against GRALIS at the 30% threshold; “—” denotes the reference. Bold indicates the best (lowest) value per column.

On Pearson faithfulness, GRALIS ranks second (0.232) behind Integrated Gradients (0.562), confirming that its region-level attributions track the direction of model-sensitivity changes, even though its maps are piecewise-constant rather than pixel-precise.

On pixel-level masking, GRALIS ranks fourth (Δ@30% = −0.032). The difference is statistically significant versus GradCAM++ (p = 0.029, Wilcoxon) but not versus GradCAM (p = 0.809); GRALIS ranks below KernelSHAP, LIME, and IG.

4.5.1 Deletion AUC—all methods, full test set

Table 5 extends the deletion AUC benchmark (Petsiuk et al., 2018) to all six methods on the complete test set (N = 1, 187), replacing the 100-image subsample previously reported for three methods only in Table 4. GRALIS ranks second (0.704 ± 0.331), trailing only IG (0.678 ± 0.222) and statistically outperforming all four remaining methods (Wilcoxon signed-rank, all p < 10⁻³³), with a gap of only +0.026 versus IG compared to gaps of +0.076–+0.100 for GradCAM, GradCAM++, KernelSHAP, and LIME. This result corroborates the Pearson faithfulness ranking (Table 9), confirming that GRALIS superpixel-level attributions identify the model's most influential regions despite operating at coarser spatial resolution than pixel-precise methods.

4.5.2 Statistical note

All Wilcoxon signed-rank p-values reported in this section are image-level and should be treated as exploratory: the patient-level split ensures that no patient appears in more than one partition, but multiple images from the same patient may share correlated activation patterns, violating the independence assumption of the test. Effect sizes (median paired differences) and cluster-robust confidence intervals by patient are left for future study.

4.5.3 Granularity-fair comparison

Since GRALIS operates on ≈ 25 superpixels, pixel-level masking penalizes its coarser spatial resolution relative to per-pixel methods. Table 10 reports a complementary protocol in which entire superpixels are masked (SLIC n_seg = 200, compactness = 10. The same segmentation is used for all methods), providing a common-granularity baseline.

Table 10. Superpixel-level masking faithfulness (N = 1, 187).

SLIC segmentation (n_seg = 200, compactness = 10) is applied once per image; all six methods share the same segmentation map. Superpixels are masked in descending order of mean attribution until the target pixel fraction is reached (fill value = 128). Δ@p%: mean ± std confidence change (more negative = better). Wilcoxon signed-rank p-values (two-sided) compare each method against GRALIS; rank is by Δ@30%. Bold indicates the best (lowest) value per column.

Under this protocol, GRALIS ranks 4 (Δ@30% = −0.045), statistically superior to both GradCAM variants (p = 0.0034 and p = 0.0016) and below LIME, IG, and KernelSHAP (all p < 10⁻⁹). The pixel-level gap to IG (0.119) shrinks to 0.019 at superpixel level, suggesting that part of IG's pixel-level advantage reflects its dense attribution granularity rather than superior region-level faithfulness.

4.6 Cross-dataset faithfulness generalization

To assess the generalization of GRALIS beyond the BreaKHis training domain, we evaluated all six methods on two held-out external datasets using the same frozen DenseNet-121 backbone and a minimal linear probe [Dropout(0.3) + Dense(1), ≈1, 025 trainable parameters, < 0.02% of total model parameters] trained separately on each target domain. For both datasets, faithfulness was measured with Deletion AUC and ROAD MoRF (Rong et al., 2022) (lower is better, 13 masking steps, 0%–100%).

4.6.1 IDC breast cancer (n = 499)

The IDC dataset (Penn Medicine/Kaggle, CC0 license) comprises 50 × 50px histological patches (249 benign, 250 malignant). The linear probe trained on 800 IDC images achieved AUC = 0.8962 (clean run; see Section 3.2). As reported in Table 11, GRALIS achieves the lowest Deletion AUC (0.189 ± 0.162) and ROAD MoRF (0.198 ± 0.137) across all six methods—ranking first on both metrics, compared to rank 2 of 6 on deletion AUC in-domain (BreaKHis). GRALIS outperforms the second-best method on Deletion AUC (GradCAM: 0.229 ± 0.189) by approximately 17%. On ROAD MoRF, GRALIS ranks first (0.198 ± 0.137) with IG as the nearest comparator (0.214 ± 0.070), an improvement of approximately 7.6%. Notably, Integrated Gradients—which ranks first in-domain (BreaKHis Del AUC = 0.678)—drops to last on IDC Deletion AUC (0.424 ± 0.173). The reason for IG's rank reversal is not established; possible factors include its pixel-level attribution mechanism, reduced discriminative activation under domain shift (probe AUC = 0.8962 vs. 0.9989 in-domain), or interaction effects specific to 50 × 50px patches.

Table 11. Cross-dataset faithfulness benchmark on the IDC dataset (Penn Medicine/Kaggle, n = 499 test patches, 50 × 50px).

Frozen DenseNet-121 backbone; linear probe AUC_IDC = 0.8962 (clean run, strict held-out evaluation set). Deletion AUC and ROAD MoRF: 13 masking steps (0%–100%); lower = better faithfulness. Best result in bold.

*PatchCamelyon/PCam (n* = 500)**. The training and evaluation protocol are described in Section 3.2 (cross-dataset linear probe, PCam). As reported in Table 12, GRALIS ranks sixth on both Deletion AUC (0.229 ± 0.167) and ROAD MoRF (0.183 ± 0.142); KernelSHAP leads Deletion AUC (0.162 ± 0.139), and IG leads ROAD MoRF (0.073 ± 0.057).

Table 12. Cross-dataset faithfulness benchmark on PatchCamelyon/PCam (Veeling et al., 2018) (n = 500 test patches, 96 × 96px, lymph node sections).

Frozen DenseNet-121 backbone; linear probe AUC_PCam = 0.8815 (clean run, strict held-out evaluation set; see Section 5.7). Deletion AUC and ROAD MoRF: 13 masking steps (0%–100%); lower = better faithfulness. All methods: N = 500, 0 failed images. Best result in bold.

A plausible explanation for the PCam rank drop is a structural granularity mismatch: with n_seg = 30 (effective ≈25 segments), each superpixel covers ≈368 native pixels (≈19 × 19px) in 96 × 96 images, while discriminative features in lymph node tissue (individual tumor cells, nuclear morphology) subtend only 5–15 px. By contrast, the same n_seg = 30 on BreaKHis (resized to 224 × 224px) yields segments covering ≈2, 000px—appropriately matched to ductal structures and stromal regions at 80–200 px scale. The full analysis and domain-adaptive n_seg recommendations are in Section 5.7.

Together, the two cross-dataset results provide a more informative characterization than a single positive result: GRALIS achieves its strongest gains where region-level anatomical coherence matters (IDC, BreaKHis), while its performance limitations on PCam are consistent with the method's superpixel granularity, though this interpretation remains to be confirmed through an explicit n_seg sensitivity analysis, and does not by itself exclude alternative explanations (domain shift, backbone specificity, masking protocol differences).

4.7 Empirical verification of design properties

Two design properties of GRALIS Algorithm 1 are verified empirically on the BreaKHis test set; full derivations are in Fanale (2026) and numerical details in Appendix 6. Exact completeness is not expected under the LIME locality kernel, whose non-uniform coalition weights break the Shapley telescoping sum; the observed empirical completeness residual (|∑iϕi-(f(x)-f(x′))|≈0.40) reflects this deliberate trade-off in exchange for LIME-style locality (see Appendix 6 for structural derivation). Variance reduction from antithetic sampling is confirmed: σ = 0.240 with antithetic pairing vs. σ = 0.339 without, yielding approximately 29% reduction in standard deviation.

4.8 GRALIS-report: generation results

The report generator was applied to the full BreaKHis test set (1,187 images) in batch mode: 1,187 reports generated, 0 failures, no external dependencies. Generation time is negligible relative to GRALIS processing (< 0.01 s per report vs. 44.8 s for attribution).

Table 13 summarizes the classification distribution. Overall, 1,175 images (98.99%) are concordant with ground truth (TP = 809, TN = 366); only 12 images (1.01%) are discordant (FP = 6, FN = 6), for which the system automatically issues a priority review alert in the Assessment and Plan fields.

Table 13. SOAP report generation: classification distribution (BreaKHis test set, 1,187 images).

Table 14 provides the breakdown of these 12 cases by optical magnification. Errors are concentrated at 400× (7 of 12, 58.3%), but the small number of discordant cases prevents reliable inference about magnification-specific error mechanisms. Three of the six false positives originate from a single patient (PT-14-21998AB); however, the significance of this observation depends on the denominator—that patient may contribute disproportionately more images than others, which is not controlled for here.

Table 14. Breakdown of 12 discordant reports by optical magnification.

FP, false positive (benign image flagged as malignant); FN, false negative (malignant image not flagged). The concentration at 400× (7/12) is a descriptive observation; no error mechanism can be reliably attributed from 12 cases.

4.8.1 Example 1—true positive (TP malignant case, 400×)

Figure 5 shows the complete system output for SOB_M_DC-14-11031-400-011.png: original image, ϕ map overlay, and top-3 superpixel highlighting. The automatically generated SOAP report reads (the Plan field contains illustrative template text and should not be interpreted as patient-specific clinical advice):

S: Breast histological sample (BreaKHis, 400x magnification). Anatomopathological ground truth: MALIGNANT.

O: Predicted diagnosis: MALIGNANT (confidence 100.0\%, high) Classification: True Positive (TP) SLIC superpixels: 25 | mean-phi: 0.6285 | Completeness: 0.3073 Top-3 regions: Region 1 (seg 17): phi = 1.0000, 1577 px (3.1\% of image) Region 2 (seg 10): phi = 0.9381, 1967 px (3.9\% of image) Region 3 (seg 8): phi = 0.9274, 1586 px (3.2\% of image) Primary attribution region occupies 3.1\% of total image area.

A: Automated report: MALIGNANT. Classifier expresses high diagnostic certainty. Attribution quality: high mean phi ($\geq$0.50), concentrated maps; low completeness proxy (good approximation of efficiency axiom). Prediction concordant with ground truth.

P: Histopathological confirmation by anatomopathologist (H&E section). WHO breast tumor staging. Multidisciplinary therapeutic planning (oncology, surgery, radiotherapy).

4.8.2 Example 2—true negative (TN benign case, 200×)

The following report was generated for SOB_B_A-14-22549AB-200-003.png, a benign adenosis case correctly classified with high confidence. In the TN setting, GRALIS produces low-magnitude, spatially diffuse attribution maps (mean ϕ < 0.20), reflecting the absence of concentrated malignancy-driving evidence; the Assessment field explicitly encodes this distributional information. As above, the Plan field contains only illustrative template text.

S: Breast histological sample (BreaKHis, 200x magnification). Anatomopathological ground truth: BENIGN.

O: Predicted diagnosis: BENIGN (confidence 97.3\%, high) Classification: True Negative (TN) SLIC superpixels: 25 | mean-phi: 0.1847 | Completeness: 0.8514 Top-3 regions: Region 1 (seg 4): phi = 0.4123, 983 px (1.9\% of image) Region 2 (seg 19): phi = 0.3801, 1204 px (2.4\% of image) Region 3 (seg 12): phi = 0.3412, 1091 px (2.2\% of image) The primary attribution region occupies 1.9\% of the total image area.

A: Automated report: BENIGN. Classifier expresses high diagnostic certainty. Attribution quality: low mean phi ($ < $0.20), diffuse maps; high completeness proxy indicates distributed, non-focal evidence---consistent with absence of dominant malignant pattern. Prediction concordant with ground truth. No priority review flag.

P: Radiological and clinical correlation recommended. Routine follow-up according to local breast screening protocol. Multidisciplinary review if clinical or radiological discordance.

4.9 Expert perception discordance study

A structured expert perception study was conducted with four anatomopathologists (P1, P2, P3, and P4) to characterize the range of expert responses to GRALIS attribution maps and GRALIS-Report SOAP outputs. The study is not intended as a clinical validation—its primary finding is methodological: the marked inter-rater variability observed illustrates substantial rater dependence in this pilot sample and cautions against treating small expert panels as explanation ground truth. Observational disclaimer: With only N = 4 raters, no inferential statistics are reported for between-rater comparisons; all figures are descriptive (means, ranges, percentages of individual responses). The marked inter-rater range (Q13: 1.47–4.73/5) is itself the primary finding: it suggests rater-dependency as a relevant factor in this domain, not a deficiency of the evaluated system, and cautions against aggregating XAI preference scores from small-N studies in histopathology. The evaluation comprised three blocks: Block A (blinded preference comparison between GRALIS and GradCAM attribution maps, 15 cases), Block B (GRALIS-Report quality assessment on 15 cases), and Block C (free annotation of regions of interest on 30 cases for inter-rater analysis). Results are summarized in Tables 15–18.

Table 15. Block A: Preference (Q1) and mean Likert scores for coherence (Q4) and utility (Q5) by AI outcome type (pooled N = 4 raters; responses per row = 4 × cases).

Method A = GRALIS, Method B = GradCAM.

4.9.1 Study design details

All four pathologists evaluated the same fixed set of cases (A01–A15 and B01–B15; Block C cases were also shared). Case identity was therefore not a confounding factor: the Q13 mean scores (1.47–4.73/5) reflect genuine between-rater variability on identical material, not between-rater-and-case-block variability. Block A was blinded with respect to method identity (Method A and Method B were anonymized); cases were labeled by classification outcome (TP/TN/FP/FN) and pathologists were not blinded to ground truth. Outcome balance per block: TP = 5, TN = 4, FP = 3, FN = 3 (Blocks A and B identical). No randomization of case order across raters was performed; cases were presented in a fixed sequence. Pathologists were not blinded to the model output in Block B (the GRALIS attribution map and report were the evaluated object). Q13 (“I would find this type of AI explanation report useful in my clinical practice”, 1–5 Likert) was assessed per case in Block B only (15 responses per rater, 60 total); individual rater means are not directly aggregable into a group estimate with N = 4.

4.9.1.1 Block A—Attribution map preference

Pooled descriptively across all four raters (60 individual responses; no consensus inference from N = 4), GRALIS was preferred in 67% of individual responses, GradCAM in 25%, and ratings were equal in 8% (Table 15). Preference for GRALIS was strongest in TN cases (81%), where benign tissue is correctly classified and the attribution map highlights normal architectural features. TP and FN cases showed more mixed signals (65% and 58% GRALIS respectively), reflecting higher inter-rater variability: P3 and P4 consistently preferred GRALIS across outcome types, while P1 and P2 showed weaker or reversed preference for certain cases. Mean Likert coherence scores (Q4) favored GRALIS over GradCAM across all outcome classes (pooled: 3.50 vs. 2.77), as did decision-support scores (Q5: 3.10 vs. 2.67).

4.9.1.2 Block B—Report quality

Mean Likert scores across map questions (Q6–Q8) and SOAP questions (Q9–Q13) are reported in Tables 16, 17. Pooled descriptively across all four raters (59–60 individual responses), TP cases achieved the highest scores on both map quality (3.47/5) and SOAP accuracy (3.65/5); FN cases scored lowest for map quality (2.56/5). Among individual questions, Q11 (second-opinion value) obtained the highest pooled mean (3.61/5), while Q7 (map region relevance) scored lowest (2.97/5). Q14 (overall assessment) was rated “Yes clearly” in 48% of cases. Inter-rater variability remained substantial. P3 rated all 15 cases “Yes clearly” (Q13 mean: 4.73/5), P4 rated 11 of 15 cases “Yes clearly” (Q13 mean: 4.27/5), while P1 rated all cases “No/Unclear” (Q13 mean: 1.47/5), and P2 showed an intermediate response (Q13 mean: 2.00/5, 3 “Yes clearly”). This divergence cautions against reading the pooled mean as a consensus clinical signal; results should be interpreted as indicating a wide range of expert assessments within a very small sample.

Table 16. Block B: Mean attribution map score (Q6–Q8), SOAP report score (Q9–Q13), and overall assessment Q14 by AI outcome type (pooled N = 4 raters; responses per row = 4 × cases).

Table 17. Block B: Per-question mean Likert scores (1–5) for each pathologist and pooled (N = 15 cases each).

P1 = Mehmet Ozcan; P2 = Abigaille; P3 = Wesley; P4 = Harshi. High inter-rater variability: P3 and P4 rated cases positively (Q13 = 4.73 and 4.27), P1 uniformly low (Q13 = 1.47).

4.9.1.3 Block C—Inter-rater annotation concordance

Inter-rater concordance between P1 and P2 in the free annotation task reached 63% under a broad definition and 57% under exact style agreement (Table 18). P3 and P4 completed all 30 cases; their Large-area and focused-circle styles were more similar to each other and closer to P2 than to P1. The main source of discordance was annotation granularity: P1 circled specific histological structures (ductal/glandular), while P2 outlined large regions. Divergence was most pronounced in TN cases (12% broad concordance), where the absence of a dominant lesion makes region selection highly subjective, and least pronounced in FN cases (83%), where pathologists with different granularity styles converged on the same structure—suggesting free-annotation tasks capture localized expert knowledge relevant for future weakly supervised localization pipelines. In the open-ended expert opinion section, a rater noted that “these systems are not intended to replace pathologists, but rather to support them during preliminary screening, prioritization, and potentially help accelerate the diagnostic workflow”—a characterization consistent with the clinical role framing in Section 3.6.3.

Table 18. Block C: Inter-rater annotation concordance by AI outcome type (N = 30 cases, N = 4 raters). Broad concordance groups Global, Large-area, and Multi-region styles as semantically compatible. P1–P2 pairwise concordance shown; P3 (Wesley) and P4 (Harshi) dominant styles listed for comparison.

5 Discussion

5.1 Two-tier reporting architecture: design rationale

From a signal-processing perspective, GRALIS-Report is a multimodal interpretation architecture that performs three successive signal transformations. The first is classical computer vision: a convolutional network maps a 2D histological image to a binary prediction. The second is attribution: GRALIS transforms that image into a sparse semantic attribution signal ϕ∈ℝns—a 25-dimensional structured representation of the model's spatial evidence. The third is cross-modal generation: a rule-based or LLM-assisted engine maps ϕ to structured natural language. Each transition is explicit, auditable, and traceable to the stored structured attribution record (F(x) CSV + raw .npy); no information is fused across modalities in a latent space, which is precisely what enables the auditability properties discussed below.

The key architectural invariant is the strict separation between the two modes: all scientific claims, benchmarking results, and faithfulness evaluations derive exclusively from Mode A (the deterministic rule-based engine). Mode A provides full reproducibility and token-level traceability, which are the properties required for audit accountability. Mode B (the optional LLM narrative layer) is included as a design component to show that the structured attribution interface F(x) can serve as a controlled gateway to language-assisted reporting without exposing raw images to the model (Cook et al., 2026); it is not evaluated quantitatively and does not contribute to any result reported in this study.

5.2 Cross-metric profile and SAL–CPT trade-off

The SAL result (0.762 ± 0.109) is consistent with the preliminary estimate (0.778 ± 0.104, Δ = 0.016), confirming implementation stability. SAL structurally favors piecewise-constant maps, so the result reflects the map type rather than independently demonstrating clinical relevance of the highlighted regions. The method comparison shows a systematic SAL–CPT trade-off: IG achieves high faithfulness but SAL = 0.150 (perceptually uninformative); LIME saturates SAL but CPT = 0.068 (near-uniform). GRALIS produces a competitive intermediate profile (SAL = 0.762, CPT = 0.018) consistent with operating in superpixel input space. Note that pixel-level CPT is structurally penalizing for piecewise-constant maps; a segment-level CPT variant would give a fairer comparison and is left for future work.

5.3 Pixel-level vs. superpixel-level ranking: granularity effect

The rank discrepancy between deletion AUC (rank 2; Table 5) and the pixel-level masking protocol (rank 4; Table 9) reflects a granularity artifact inherent to piecewise-constant attribution maps. Deletion AUC removes pixels in descending attribution order: GRALIS assigns uniform high scores to entire superpixels, so its highest-ranked pixels form spatially coherent blocks that efficiently erode model confidence. The pixel-masking Δ@p% protocol removes a fixed fraction p of pixels by attribution rank; at p = 10%–20%, the discrete superpixel structure causes GRALIS to remove either no complete segment (if the top-p% fraction does not span a full superpixel) or entire ones at a threshold jump—producing coarser descent in confidence than the smooth per-pixel removal of IG or SHAP. The superpixel-level masking protocol (Table 10) corrects for this: the remaining gap (GRALIS rank 4, Δ@30% = −0.045 vs. IG −0.064) shrinks considerably from the pixel-level gap (−0.032 vs. −0.120), confirming that most of IG's advantage under pixel masking derives from granularity rather than superior region identification. The class-conditional superpixel deletion test further supports this reading: top-3 superpixels move confidence in the predicted direction in 96% of cases, the strongest signal among all evaluated protocols.

5.4 Per-magnification ExpiScore analysis

The BreaKHis test set spans four optical magnifications (40×, 100×, 200×, 400×). GRALIS is applied identically across all magnifications without magnification-specific parameters; the term multiscale in the paper title refers to this cross-magnification coverage. Table 19 reports ExpiScore components by magnification on the stratified test set.

Table 19. Per-magnification ExpiScore breakdown for GRALIS on the BreaKHis test set (N = 1, 187 images, patient-level split).

SAL, salience concentration; CPT, compactness; CSC, cross-method structural consistency; ExpiScore = (SAL+CPT+CSC)/3. Standard deviations computed across images within each magnification subset.

Kruskal–Wallis, permutation ANOVA, and BIC-based Bayes Factor tests all confirm magnification independence: no component shows a significant cross-magnification difference (BF₀₁ > 15, 000 for every metric, decisive on the Jeffreys scale (Jeffreys, 1961; Wagenmakers, 2007; Kruskal and Wallis, 1952; Anderson, 2001)). CPT remained consistent across all magnifications (range 0.018–0.019; Figure 6), suggesting that SLIC superpixel segmentation adapts to varying tissue scales.

5.5 Class-conditional faithfulness

Superpixel deletion testing shows that GRALIS faithfulness is class-symmetric: for malignant images it identifies pro-malignant superpixels (96% positive, K = 3); for benign images it identifies anti-malignant superpixels (96% negative, K = 3). This behavior, initially counter-intuitive, is theoretically correct and clinically relevant: salient regions for a benign prediction (e.g., well-differentiated glands, absence of mitoses) are distinct from those for a malignant prediction (e.g., nuclear pleomorphism, stromal invasion).

5.6 Sensitivity analysis: selection of k and baseline

Attribution score stability is evaluated across 15 configurations (k ∈ {10, 20, 50, 100, 200} × three baselines: blur, mean, zero) on a 100-image subsample. Blur baseline shows earliest convergence (k = 20, Δ < 0.0004); zero baseline converges at k = 50 and is adopted for main results to ensure comparability with Integrated Gradients. Figure 7 summarizes all configurations.

5.7 Limitations

5.7.1 Dataset scope

Primary results are on BreaKHis; cross-dataset evaluations on IDC (n = 499) and PCam (n = 500) provide cross-dataset transfer evidence across three datasets and two tissue types (breast histology, lymph node). All cross-dataset experiments use a frozen backbone with a minimal linear probe (1,025 parameters, < 0.02% of total model weights). Generalization to multi-scanner, multi-institution cohorts still requires independent future replication.

5.7.2 Cross-dataset granularity mismatch (PCam)

On PatchCamelyon, GRALIS ranks sixth on both Deletion AUC and ROAD MoRF (Table 12). With n_seg = 30 applied to 96 × 96px native images, each segment covers ≈368 pixels (≈19 × 19px), while discriminative structures in PCam (individual tumor cells, lymphocyte clusters) subtend only 5–15 px at this acquisition scale (Veeling et al., 2018). The superpixel grid is therefore too coarse to isolate individual cellular structures, conflating discriminative and non-discriminative tissue within each segment and inflating attribution variance. Additionally, SLIC compactness = 50 produces spatially regular (near-rectangular) segments appropriate for boundary-defined anatomical features (Achanta et al., 2012); PCam discrimination is texture-based (lymphocyte density gradients), for which lower compactness values (10–20) would be more appropriate. Domain-adaptive n_seg configuration (n_seg ∈ {80, 120} with compactness ∈ {10, 20}) for patch-based texture datasets is left for future work. This result demonstrates that GRALIS superpixel advantages are domain-specific (strongest where region-level anatomical coherence matches the feature scale) and that the performance profile is consistent with the method's segmentation parameters, though a dedicated n_seg ablation would be required to confirm this interpretation. Notably, the poor average rank does not reflect uniform failure: GRALIS achieves a mean per-image Del-AUC rank of 4.57 (rank distribution R1–R6: 13, 30, 49, 122, 136, 149 images out of 500) with a coefficient of variation of 73% (CV = σ/μ), consistent with a situational mismatch rather than a systematic attribution failure. On 34% of images (170/500) GRALIS falls within 0.05 of the best method's Deletion AUC, and achieves rank 1 on 13 images (2.6%); when the discriminative region coincides with the SLIC superpixel grid, GRALIS performs competitively. LIME—also a superpixel-based method using SLIC—ranks fifth on Deletion AUC (mean 0.219, rank distribution analogous to GRALIS), providing indirect support for the granularity-mismatch interpretation.

5.7.3 PCam probe evaluation set—corrected run

An initial probe run used the 500-image official PCam test split also as the validation set for early stopping, introducing indirect test-set exposure. A corrected run was performed with a strict held-out protocol: the 800 official PCam training images were divided 80/20 (640 train, 160 val internally) and the 500 evaluation images were held out entirely during training and epoch selection. Probe AUC changed by Δ = −0.006 (0.8876 → 0.8815). All faithfulness rankings are invariant to the protocol correction: GRALIS ranks sixth on both Deletion AUC and ROAD MoRF in both runs; KernelSHAP is rank 1 on Deletion AUC in both runs. The numbers reported in Table 12 and throughout this research reflect the corrected protocol.

5.7.4 Human evaluation

The expert perception study (N = 4, 60 cases) is a methodological pilot; its primary contribution is documenting marked inter-rater variability (Q13 means: 1.47–4.73/5) as a calibration signal for XAI evaluation.

5.7.5 Benchmark coverage

ROAD is reported for GRALIS, IG, and GradCAM; additional metrics and methods are planned. The ExpiScore conflict of interest is declared in Section 2.5.

6 Conclusion

We have presented GRALIS-Report, a research prototype for computer-aided explainability in breast histology, combining attribution maps grounded in a principled theoretical framework with a downstream SOAP report generation system. On the supplementary ExpiScore_*s* profile, GRALIS achieves SAL = 0.762 ± 0.109 and ranks second on CPT and deletion AUC (Table 4); detailed rankings are author-defined and should be read alongside independent metrics. The framework shows class-conditional faithfulness (96% consistency with K = 3 superpixels on malignant images). Independent pixel-level deletion AUC (Petsiuk et al., 2018) on the full test set (N = 1, 187) places GRALIS (0.704 ± 0.331, rank 2 of 6) directly below IG (0.678 ± 0.222, rank 1) and above all GradCAM and superpixel-surrogate methods (Table 5), consistent with the trade-off between superpixel-level coherence and pixel-level precision. ROAD (Rong et al., 2022) evaluation on the same test set confirms this ranking (MoRF AUC: IG 0.677, GRALIS 0.704, GradCAM 0.755; lower is better) and reveals that GRALIS achieves the largest MoRF–LeRF discrimination gap (0.140 vs. 0.098 for IG and 0.030 for GradCAM), indicating the strongest separation between relevant and irrelevant regions among the three evaluated methods. The GRALIS-Report system generated 1,187/1,187 syntactically complete SOAP reports with no execution failures; 1,175 (98.99%) corresponded to correct classifier predictions, with 12 discordant cases identified retrospectively using test-set labels, operating fully offline. Cross-dataset experiments on two held-out external datasets further characterize the trade-off. On IDC Breast Cancer (n = 499, probe AUC = 0.8962, clean run), GRALIS ranks first on both Deletion AUC (0.189) and ROAD MoRF (0.198), providing evidence that the attribution mechanism can retain faithfulness under this cross-dataset transfer setting (Table 11). On PatchCamelyon (n = 500, probe AUC = 0.8815), GRALIS ranks sixth—a result plausibly explained by n_seg = 30 being too coarse for the 5–15 px scale of lymph node features (granularity mismatch hypothesis), and is reported in full transparency (Table 12, Section 5.7). These contrasting results are more informative than a single positive outcome: GRALIS faithfulness advantages are specific to settings where superpixel granularity matches the target feature scale. The companion paper (Fanale, 2026) provides the theoretical basis for the design of the GRALIS attribution method; the present study reports exclusively empirical findings.

The present study should be read as an engineering and exploratory framework paper: it demonstrates the feasibility of a cross-modal attribution-to-report pipeline, and honestly reports its limitations — including the self-referential benchmarking structure and the rater-dependent clinical utility signal from the discordance study. These limitations define the agenda for future independent validation work.

Future work is organized around six concrete directions.

• Faithfulness benchmarking. ROAD (Rong et al., 2022) has been computed for GRALIS, IG, and GradCAM on the full BreaKHis test set; results are reported in Section 4.5. Remaining gaps: ROAD for KernelSHAP, LIME, and GradCAM++; insertion AUC for all methods; sanity checks (Adebayo et al., 2018); and pointing-game localization once bounding-box annotations are obtained for BreaKHis.
• External dataset validation. Cross-dataset evaluations on IDC (n = 499, breast histology) and PatchCamelyon (n = 500, lymph node) are reported in Section 4.6. GRALIS ranks first on IDC (both metrics) and sixth on PCam (granularity mismatch, Section 5.7). Full independent replication on a multi-scanner, multi-institution cohort remains necessary before any clinical generalization claim can be made. Planned target datasets include TCGA-BRCA and at least one European cohort to test stain normalization robustness across H&E preparation protocols. Domain-adaptive n_seg for patch-based texture datasets (PCam, CIFAR-histology) is a natural and well-motivated extension.
• Whole-slide image (WSI) extension. Current BreaKHis images are pre-cropped tiles; scaling GRALIS to WSI-level analysis via sliding-window or patch-based tiling is a necessary step for clinical adoption.
• Texture-aware superpixel segmentation. Pathologist feedback indicated that texture recognition and attribution rendering did not yet meet clinical expectations; texture-aware SLIC strategies better aligned with H&E inspection criteria should be explored.
• Artefact pre-screening. Several images presented acquisition artifacts (e.g., air pockets) that may have contributed to false-positive outputs; an explicit slide-quality pre-screening step and scanner-robustness evaluation are needed.
• Scaled human evaluation and Mode B assessment. The current four-pathologist perception study reveals high inter-rater variability, underscoring the need for a properly powered study (pre-registered protocol, independent raters, N ≥ 100 cases) to resolve the discordance and convert this preliminary pilot into actionable clinical evidence. Quantitative evaluation of Mode B narrative quality (terminological adequacy, clinical utility on structured Likert scales) will be included in the extended version.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. For language editing. All scientific content, experimental results, data analyses, figures, and conclusions are the sole responsibility of the author.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Dasara Shullani, University of Florence, Italy

Reviewed by: Xi Xiao, University of Alabama at Birmingham, United States

Nauman Ullah Gilal, Hamad Bin Khalifa University, Qatar