Explainable artificial intelligence for the analysis of histopathological images of breast cancer: methods, interpretability and emerging directions

1 Introduction

Digital histopathology represents one of the most challenging domains in medical image analysis due to the complexity, scale and variability of the visual information involved (Foucart et al., 2022). Whole Slide Images (WSIs) of breast tissue are characterized by extremely high spatial resolution and substantial color variability induced by staining protocols (Modi et al., 2024) as illustrated in Figures 1, 2, and morphological structures organized across multiple spatial scales, from cellular detail to global tissue architecture (Nguyen et al., 2017). The analysis of such images is inherently demanding and remains affected by inter- and intra-observer variability in manual assessment.

Recent advances in machine learning and deep learning have enabled the development of automated and semi-automated systems capable of analyzing WSIs with high predictive performance. Convolutional neural networks, attention-based architectures and transformer-based models have demonstrated strong results in tasks such as classification, segmentation and risk stratification (Unger and Kather, 2024; Springenberg et al., 2022; Edafetanure-Ibeh et al., 2025; Mondol et al., 2025). However, the increasing complexity of these models has also amplified concerns regarding their transparency and reliability, particularly in sensitive application contexts such as medical imaging.

A key limitation of many existing approaches lies in their black-box nature. While accurate predictions can be obtained, the lack of clear and verifiable explanations makes it difficult to assess whether model decisions are driven by meaningful morphological patterns or by confounding factors and acquisition artefacts (Wang et al., 2024; Dawood et al., 2024). This issue has been widely recognized as a barrier to trust, validation and broader adoption of deep learning systems in histopathological practice (Abdullakutty et al., 2024b).

Explainable Artificial Intelligence (XAI) has emerged as a response to these concerns, aiming to provide insights into the internal mechanisms and decision processes of complex models. In the context of histopathological image analysis, explainability plays a crucial role in linking algorithmic outputs to interpretable visual patterns, thereby supporting model validation and critical assessment (Mondol et al., 2025; Somani et al., 2024). These explanations are typically expressed as intuitive spatial relevance maps that highlight image regions contributing to a model’s prediction, but they do not capture the full complexity of the model’s internal representations (Figure 2).

Despite growing interest in XAI, existing approaches remain fragmented and heterogeneous. Different methods provide explanations at different levels of representation, ranging from spatial relevance maps to feature attribution scores and semantic concepts (Mersha et al., 2024). As a result, explanations are often difficult to compare, integrate and evaluate in a systematic manner, particularly when dealing with multiscale image data such as WSIs.

This work is a Mini Review Article and does not present new experimental results or original model evaluations. Instead, it critically analyzes and synthesizes existing literature on explainable artificial intelligence (XAI) techniques applied to breast cancer histopathological image analysis, with a specific focus on interpretability from an image-processing and signal-level perspective. The review provides a structured and critical examination of the main XAI approaches, emphasizing interpretability at the level of visual representations. Rather than offering a purely descriptive overview, the paper evaluates the strengths and limitations of current methods, highlights recurring methodological challenges, and discusses emerging research directions aimed at improving the robustness, multiscale coherence, and practical usefulness of explanations within advanced image analysis pipelines.

2 Image processing in breast cancer histopathological images

As a review work, this section summarizes commonly adopted pipelines and model families from the literature.

Whole Slide Image (WSI) can be regarded as ultra-high-resolution two-dimensional signals characterized by a hierarchical multiscale structure (Mondol et al., 2025; Sharma, 2025). Relevant diagnostic information is distributed across multiple levels, including local cellular textures, nuclear patterns, glandular organization and global stromal structure (Nguyen et al., 2017). This multiscale nature necessitates the use of tiling, downsampling and feature aggregation strategies. Early image processing approaches relied on hand-crafted features (e.g., Haralick textures, Local Binary Patterns, Gabor filters), whereas deep learning models learn latent representations directly from data (Unger and Kather, 2024). Convolutional Neural Networks, Vision Transformers and graph-based models are now widely used to capture complex spatial and contextual relationships (Mondol et al., 2025; Springenberg et al., 2022; Edafetanure-Ibeh et al., 2025).

Despite their high performance, DL models tend to learn correlations that are difficult to interpret. In the absence of explainability mechanisms, it is challenging to verify whether the model is truly exploiting histopathologically relevant patterns or confounding signals.

3 Explainable artificial intelligence: techniques for histopathological image analysis

In image processing, explainability is closely linked to the ability to map model outputs onto spatial regions and signal features of the input. An effective explanation should be spatially coherent, stable and faithful to the actual behavior of the model (Somani et al., 2024; Graziani et al., 2021). Key principles of XAI applied to histopathological images include transparency, fidelity, stability, multiscale coherence and domain-relevant interpretability (Somani et al., 2024; Mersha et al., 2024).

While Figure 2 provides a visual illustration of saliency-based explanations, the full spectrum of XAI techniques reviewed in this work—including feature attribution, perturbation-based, concept-based, prototype-based, and intrinsically interpretable approaches—is systematically discussed and comparatively analyzed in the following sections and summarized in Table 1.

TABLE 1. XAI techniques: methods, advantages and limitations.

3.1 Saliency- and gradient-based methods

Saliency and Gradient-based methods represent one of the most widely used classes of XAI in image processing.

Saliency-based methods operate on intermediate feature representations of deep neural networks, typically convolutional layers, to identify spatial regions that most influence model predictions (Peta and Koppu, 2023; IEEE Access, 2024; Krishna et al., 2023).

A widely used approach is Grad-CAM, which computes the importance weight of each feature map A k for a class c as:

α k c = 1 Z ∑ i ∑ j ∂ y c ∂ A i j k

The final localization map is then obtained as:

L Grad ‐ CAM c = ReLU ∑ k α k c A k

This formulation leverages gradients as indicators of sensitivity, effectively projecting class-specific importance back onto spatial feature maps.

In histopathology, these maps enable the localization of tumor-relevant regions and allow verification of whether the model focuses on morphologically relevant structures, such as atypical nuclei, regions of high cellular density or invasive fronts (Mondol et al., 2025; Peta and Koppu, 2023). Across commonly used WSI and patch-based datasets, these methods have shown effectiveness in highlighting diagnostically relevant regions, although they often produce coarse and unstable explanations. From a signal processing perspective, these techniques act as spatial relevance filters but suffer from known limitations in terms of stability, resolution and noise sensitivity (Graziani et al., 2021; Sharma et al., 2024).

3.2 Feature attribution methods

Feature attribution approaches, including SHAP and LIME, estimate the contribution of individual features or image regions to the model output (Graziani et al., 2021; Peta and Koppu, 2023; Hussein et al., 2025). When applied to histopathological images, they enable quantification of the relative importance of specific visual patterns or signal components.

Integrated Gradients computes feature importance along a continuous path from a baseline input x ′ to the actual input x:

I G i x = x i − x i ′ ∫ 0 1 ∂ F x ′ + α x − x ′ ∂ x i d α

Similarly, SHAP values are based on cooperative game theory and compute the marginal contribution of each feature:

ϕ i = ∑ S ⊆ N ∖ i ∣ S ∣ ! ∣ N ∣ − ∣ S ∣ − 1 ! ∣ N ∣ ! f S ∪ i − f S

These methods provide additive and theoretically justified explanations, often satisfying properties such as completeness and consistency.

Although these methods provide formally grounded local explanations, when applied to breast cancer histopathology datasets they offer more stable feature importance estimates, but they are computationally expensive and often difficult to interpret in high-dimensional domains such as WSIs (Graziani et al., 2021). Their fidelity also strongly depends on the quality of the local model approximation and baseline selection.

3.3 Concept- and prototype-based approaches

Concept Activation Vectors (TCAV) represent a significant advancement, as they allow evaluation of the influence of interpretable semantic concepts (e.g., “tumor invasion” or “benign stroma”) on model decisions. This approach reduces the gap between numerical representations and human interpretation.

A representative approach is TCAV (Testing with Concept Activation Vectors), which measures how sensitive a model’s prediction is to a predefined concept direction v C in latent space:

TCAV C , k = 1 ∣ X k ∣ ∑ x ∈ X k 1 ∇ f k x · v C > 0

This formulation evaluates whether increasing the presence of a concept (e.g., “cellular irregularity”) increases the likelihood of a given class.

Similarly, case-based reasoning and prototype-based models explain predictions by comparing inputs with representative training examples (Mersha et al., 2024) such as nuclear atypia or tissue structure. In studies on histopathological datasets, concept-based approaches such as TCAV enable quantification of clinically meaningful concepts; however their effectiveness strongly depend on dataset quality, coverage, subjectivity and concept definitions.

3.4 Intrinsically interpretable architectures

Beyond post hoc explanations, recent literature highlights growing interest in explainable-by-design architectures. Explicit attention mechanisms, interpretable Graph Neural Networks and multimodal models with attention weights integrate explainability directly into the learning process (Mondol et al., 2025; Edafetanure-Ibeh et al., 2025; Zhang et al., 2025).

Examples include:

• Attention-based models, where predictions are weighted by learned attention scores:
• Prototype-based networks, where classification is based on similarity to learned prototypes:
• Graph Neural Networks (GNNs), where nodes (e.g., cells or regions) and edges encode tissue relationships, and predictions emerge from message passing: These architectures embed interpretability at different levels, including spatial attention, similarity reasoning, and structural relationships.

From an image processing perspective, these models offer more stable and signal-consistent explanations, reducing the risk of artefactual relevance maps (Mondol et al., 2025; Krishna et al., 2023). In particular, graph-based and attention-based architectures have demonstrated improved interpretability and performance in WSI analysis tasks, although they introduce increased computational and modeling complexity.

4 Integrating machine learning and XAI in image processing pipelines

Integrating ML and XAI is a key step in histopathological pipelines. While traditional systems focus primarily on predictive performance, this approach is insufficient in high-stakes medical contexts without an adequate understanding of model behaviour.

Explainability-aware pipelines integrate XAI modules at different stages, including pre-processing, feature extraction, inference and output validation (Abdullakutty et al., 2024b; Wang et al., 2024). From a signal processing viewpoint, XAI can be interpreted as a secondary signal analysis layer that highlights spatial, structural or frequency components influencing system responses.

Histopathological images contain information distributed across multiple spatial scales. ML–XAI integration enables assessment of explanation consistency across magnification levels (Mondol et al., 2025; Sharma, 2025). In multimodal models combining histopathological images with clinical or radiological data, XAI also allows evaluation of the relative contribution of each modality (Abdullakutty et al., 2024a).

Explainable Artificial Intelligence techniques developed for histopathological image analysis are increasingly being investigated across other breast cancer imaging modalities, including mammography, ultrasound and magnetic resonance imaging (MRI). Although these modalities differ substantially in acquisition physics, spatial resolution and signal characteristics, they share common challenges related to model transparency, trust and clinical validation.

In mammography, XAI methods are typically applied to low-contrast, large-field images where diagnostic cues are often subtle and globally distributed. Saliency-based techniques such as Grad-CAM and attention maps are commonly used to highlight suspicious regions, such as microcalcifications or mass boundaries, supporting radiologists in correlating algorithmic outputs with known radiological patterns. However, the limited spatial resolution of explanations and sensitivity to noise remain critical issues.

Ultrasound imaging presents distinct challenges due to speckle noise, operator dependence and high texture variability. In this context, feature attribution methods and prototype-based explanations have shown potential in identifying discriminative texture patterns while maintaining interpretability. Recent studies suggest that explainability plays a particularly important role in ultrasound, where diagnostic uncertainty is inherently higher.

MRI-based breast cancer analysis introduces an additional temporal and multiparametric dimension. Explainability techniques have been extended to highlight not only spatial regions but also temporal signal dynamics and modality-specific contributions (Zhang et al., 2025). Attention-based architectures and multimodal XAI approaches are increasingly explored to disentangle the relative influence of different MRI sequences.

Overall, while most XAI methods are modality-agnostic at a conceptual level, their practical behavior and limitations are strongly influenced by the underlying signal properties. This highlights the need for modality-aware explainability strategies and for systematic evaluation of explanation robustness across heterogeneous imaging domains.

Key benefits include signal validation, bias detection, robustness and generalization analysis, and architecture optimization guided by explanatory evidence.

5 Open challenges and emerging directions

Despite substantial progress in Explainable Artificial Intelligence applied to histopathological image analysis, several open challenges continue to hinder its systematic adoption in advanced image analysis pipelines. These challenges are not limited to methodological aspects but are closely related to the intrinsic complexity of histopathological images and the way information is represented and aggregated by deep learning models.

A primary limitation concerns the fragmentation of explanatory paradigms. Saliency- and gradient-based approaches mainly provide spatially localized explanations, while feature attribution methods operate in latent representation spaces and concept-based techniques aim to introduce semantic abstractions (Somani et al., 2024; Mersha et al., 2024). As reported across multiple recent studies, these explanatory layers are typically analyzed in isolation, resulting in explanations that are difficult to compare and integrate.

This fragmentation limits understanding of how information propagates from raw data to predictions. As a consequence, explanations may appear visually plausible while remaining weakly connected to the transformations that actually drive model decisions.

Another challenge concerns explanation stability. Numerous works have shown that minor variations in input images—such as changes in staining intensity, tiling configuration or compression artefacts—can lead to substantially different relevance patterns (Springenberg et al., 2022; Graziani et al., 2021). Such sensitivity undermines confidence in the explanatory output, particularly in clinical contexts where consistency is essential. Despite its importance, stability is still rarely treated as a first-class evaluation criterion in XAI studies on histopathological imaging.

Closely related is the issue of multiscale coherence. Diagnostic information in breast cancer histopathology is distributed across multiple spatial resolutions, ranging from cellular morphology to tissue-level organization. However, many explanation methods generate relevance maps at a single scale, often corresponding to the deepest layers of the network (Mondol et al., 2025). Recent studies implicitly highlight that explanations lacking consistency across magnification levels risk misrepresenting the actual decision-making process of the model.

A further open problem concerns the absence of standardized evaluation frameworks for explainability. While predictive performance is assessed using well-established quantitative metrics, the quality of explanations is often evaluated qualitatively or through proxy measures that capture only limited aspects of interpretability (Kunhoth et al., 2024; Somani et al., 2024). There is currently no widely adopted framework capable of jointly assessing spatial localization accuracy, stability under perturbations, coherence across scales and faithfulness to the underlying model.

In response to these limitations, an emerging research direction advocates for composite and multidimensional views of explainability. Rather than relying on a single explanatory technique, several recent works suggest combining complementary explanations to capture different properties of model behavior (Abdullakutty et al., 2024b; Mersha et al., 2024). This perspective aligns with established practices in image analysis, where complex systems are characterized through multiple criteria rather than a single descriptor.

Addressing these challenges requires moving toward integrated and evaluation-driven explainability approaches.

6 Patient-centered explainability and communication

While clinicians and pathologists require detailed, signal-level explanations, patients benefit from simplified, high-level representations that convey rationale without technical complexity.

Patient-centered XAI approaches typically rely on simplified visual explanations, narrative summaries or counterfactual scenarios that answer intuitive “what-if” questions (Mersha et al., 2024). For example, saliency maps can be adapted to show patients which image regions influenced a diagnosis, while feature attribution methods can be translated into qualitative descriptions of contributing factors.

From a signal processing perspective, this implies a progressive abstraction of explanations, moving from pixel- or feature-level relevance towards semantically meaningful concepts (Somani et al., 2024). Such abstraction is essential to avoid cognitive overload and to support informed consent and trust.

Although patient-facing explainability remains underexplored in histopathological imaging, its integration represents an important future direction, particularly as AI-based diagnostic systems become more directly involved in clinical decision pathways.

7 Discussion

This discussion critically examines the XAI approaches within the context of histopathological image analysis, focusing on conceptual implications rather than methodological descriptions. The analysis highlights several issues that emerge when explainability methods are applied to highly complex and multiscale visual data.

A central issue concerns the mismatch between representational complexity and explanatory simplicity. Deep learning models used for Whole Slide Image analysis learn hierarchical representations that integrate information across scales, yet many explanation techniques reduce this complexity to simplified spatial relevance maps. While such visualizations are intuitive, they often fail to convey how different components of the representation interact to produce a decision.

Multiple studies suggest that visually convincing explanations do not necessarily correspond to diagnostically meaningful structures (Dawood et al., 2024; Graziani et al., 2021). This limitation is particularly evident in pixel-centric approaches, where relevance is assigned locally without adequately accounting for spatial organization and contextual dependencies that are critical in histopathology.

Another key aspect is explanation robustness. Reliable explanations should exhibit consistent behavior under bounded variations of the input. However, a growing body of work reports significant variability in saliency maps and attribution scores in response to minor perturbations (Springenberg et al., 2022; Graziani et al., 2021; Sharma et al., 2024). This instability raises concerns about the reliability of explanations, especially in high-stakes medical applications where interpretability is expected to support trust and accountability.

The inherently multiscale nature of histopathological images further complicates interpretability. Although deep learning models implicitly exploit hierarchical features, current XAI methods rarely expose how relevance is distributed across different levels of abstraction. Explanations that are informative at the cellular scale may overlook tissue-level organization, while those highlighting global patterns may obscure local morphological details. This disconnect limits the ability of domain experts to reconcile explanations across magnification levels.

These observations suggest that explainability should be reconsidered as a multiresolution analysis problem, in which relevance is examined across spatial scales rather than at a single resolution. Such a view provides a more faithful representation of how information is structured and utilized by modern image analysis models.

The discussion also reveals a persistent gap between algorithmic explanations and domain semantics; the explanations may not align with histopathological concepts that are meaningful to experts. Concept-based and prototype-driven approaches partially address this issue, but they introduce additional challenges related to concept definition, annotation effort and variability in expert interpretation.

Overall, explainability cannot be captured by a single method or metric. Properties such as spatial localization, stability, multiscale coherence and faithfulness represent complementary dimensions of interpretability. Evaluating these dimensions in isolation provides only a partial view of model behavior.

In this light, a multidimensional perspective on explainability appears both necessary and timely. Rather than treating explainability as an auxiliary post-processing step, it should be integrated into the design and evaluation of image analysis pipelines, informing model development and validation in a systematic manner.

8 Conclusion

This review has examined the current landscape of Explainable Artificial Intelligence techniques applied to the analysis of breast cancer histopathological images, with particular attention to how explanations relate to the underlying visual representations learned by modern image analysis models. While deep learning methods have achieved high predictive performance, the analysis highlights that explainability remains a critical and unresolved aspect of their deployment in sensitive medical contexts.

The reviewed literature indicates that existing XAI approaches provide complementary but often fragmented insights into model behavior. Saliency-based methods offer intuitive spatial visualizations, feature attribution techniques quantify local contributions, and concept-based approaches attempt to bridge numerical representations and semantic interpretation. However, each of these methods captures only a subset of the properties required for a comprehensive understanding of model decisions.

Recurring limitations include instability under minor input perturbations, limited coherence across spatial scales and the absence of shared evaluation criteria for assessing explanation quality. These issues complicate both the comparison of different models and the interpretation of their outputs by domain experts.

Taken together, the evidence suggests that explainability cannot be adequately addressed through isolated methods or single evaluation dimensions. Instead, a multidimensional perspective is required, in which different explanatory properties—such as spatial localization, robustness, multiscale consistency and faithfulness—are considered jointly.

Without proposing formal metrics or definitive frameworks, this review highlights an emerging shift towards more integrated views of explainability, in which multiple explanatory signals are combined to provide a more reliable characterization of model behavior. Such a perspective appears particularly relevant for complex image analysis tasks involving high-resolution, multiscale data.

In conclusion, integrating explainability into histopathological pipelines is essential for model design, evaluation and validation. Advancing towards more transparent, robust and interpretable systems will be key to ensuring the scientific credibility and practical impact of deep learning methods in this domain.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

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: Steven Fernandes, Creighton University, United States

Reviewed by: Prashant Narayankar, KLE Technological University, India