Multi-stage generative enhancement and foundation model framework for automated dental implant identification: assessing the preprocessing paradox

Artificial intelligence-based image analysis is transforming medical diagnosis; however, its clinical usefulness is limited by inherent noise and the quality of available medical datasets. Recent advances in generative AI promise to improve image quality through denoising, enhanced resolution, and more. This study applies the Segment Anything Model 2 (SAM2) and the Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN) to dental radiographic datasets. A dataset of 6,413 radiographs across 35 dental implant systems was used to explore the pre-processing paradox, the relationship between human visual clarity and machine-learning accuracy. First, YOLOv11 detected and cropped implant areas, establishing the baseline for two classification models: ConvNeXt and ResNet50. These were compared with three experimental variants: SAM2 segmentation, ESRGAN super-resolution, and a fully integrated YOLO+SAM2 + ESRGAN pipeline. Performance was assessed using accuracy, F1-score, PSNR, and SSIM. In the baseline YOLO+CNN approach, the accuracy of ConvNeXt (83%) outperformed that of ResNet50 (79%). Adding SAM2 segmentation lowered performance (ConvNeXt: 81%, ResNet50: 76%), suggesting that removing the bone-implant interface eliminates important contextual signals. Both ConvNeXt and ResNet maintained their average accuracies (83 and 80%) when ESRGAN was applied to raw and cropped images. The largest difference appeared in the fully integrated pipeline, where ConvNeXt achieved higher accuracy (82%) compared to a significant decline in ResNet performance (64%). The study confirms the pre-processing paradox. While SAM2 and ESRGAN enhance visual clarity, the artifacts they introduce can reduce ResNet’s accuracy. The YOLOv11-CNN approach is optimal, and ConvNeXt is recommended for its robustness to image preprocessing. However, the proposed preprocessing pipeline did not significantly improve accuracy.

Identification of dental implants is a critical aspect of forensic and clinical dentistry, especially given the maturing global market with multiple unique implant systems. Millions of patients worldwide have had prosthetic components implanted, often decades ago, and there is no universal tracking system for the brands used; hence, the development of digital tools becomes necessary (Kim et al., 2020).

The global popularity of dental implants has increased tremendously. This boom has led to a major clinical problem, frequently referred to as the “orphan implant” phenomenon (Taşdemir et al., 2025; Sharma et al., 2020). Patients visit the dentist years after implant placement with biologic and mechanical complications, and at that time, they may not have their implant records. Hence, dental practitioners must rely on implant morphology evaluation using two- dimensional (2D) radiographs to determine the brand and prosthetic platform required for repair (Takahashi et al., 2020). This method leads to errors because most implant systems share similar overall geometry, making differentiation difficult (Hadj Saïd et al., 2020).

The integration of artificial intelligence into medical image analysis has reshaped diagnostic workflows by enabling the automatic detection of complex anatomic features. In dentistry, deep learning methodologies are gaining ground as a means to overcome the limitations of subjective human judgment in complex radiological examinations (Ali et al., 2025; Stephan et al., 2024). The study explores the integration of computer vision and clinical implantology. It presents a multi-phase deep learning pipeline that addresses the challenging task of interpreting implant brands from radiographs through highly accurate segmentation and image enhancement.

While deep learning has enabled automated classification with Convolutional Neural Networks (CNNs), a major gap remains in end-to-end clinical applications. Initial models typically focused on perfect, pre- cropped images and thus did not account for the noise in unprocessed radiographs (Balel et al., 2025). Robust pipelines are missing that would link state-of-the-art detection (YOLOv11) with top-notch segmentation and enhancement. Besides that, there is a preprocessing paradox: on the one hand, foundation models such as SAM 2 and generative models like ESRGAN can make images appear clearer to humans. However, it remains unclear whether these changes preserve the contextual signals, e.g., the bone-implant interface, necessary for neural networks (Dong et al., 2026; Masluha and Azhar, 2025). Moreover, there is insufficient data on how transformer- inspired models, such as ConvNeXt, respond to generative artifacts compared with traditional baselines, such as ResNet (Liu et al., 2022). Recently, Lashaki et al. (2025) proposed an optimized CNN-based framework for radiographic image preprocessing. The study highlights the influence of preprocessing techniques on feature extraction and on CNN classification performance. The findings underscore the importance of image quality and indicate that preprocessing strategies will substantially influence deep learning performance.

Global advances have also fueled the rapid growth of dental biomaterials, transforming implantology. Calcium phosphate biomaterials, such as hydroxyapatite, are widely used for their biocompatibility and ability to promote bone regeneration around implants. These materials play a prominent role in enhancing osseointegration, preserving alveolar bone, and improving long-term implant stability. These advances help implants last longer and may necessitate future repair and maintenance, underscoring the need for implant identification (Kim et al., 2020).

This study seeks to narrow these gaps by benchmarking a modular, multistage pipeline for dental implant identification. The performance of a YOLOv11-CNN ensemble is evaluated for automatic implant detection and brand classification. Then, the influence of the SAM2-based segmentation is assessed in isolating anatomical structures by removing bone-implant features, which may lead to the loss of vital contextual features. Besides, the study examines the influence of the ESRGAN super-resolution technique on the recovery of micro-geometric details and on model performance. Lastly, a head-to-head comparison between ConvNeXt and ResNet is conducted to determine which model is more robust to the artifacts introduced by a segmented and super-resolved clinical pipeline. The principal novelties of the work are summarized as follows:

1. A fully automated, multi-stage framework that integrates YOLOv11, SAM2, Real-ESRGAN, and CNN classifiers for dental implant identification.

2. A detailed investigation of the Preprocessing Paradox and image enhancement, along with a comprehensive comparison between conventional (ResNet-50) and modern (ConvNeXt-Tiny) architectures under segmentation and generative enhancement.

3. Evaluation on one of the largest dental implant datasets from India reported to date (6,413 radiographs representing 35 implant systems).

This section provides detailed technical guidance for implementing the proposed multi-stage deep learning pipeline. The study employs a layered experimental design to evaluate a basic detection-classification framework, as well as advanced streams that integrate foundation-model-based segmentation and generative super-resolution. Figure 1 presents the flow diagram for the proposed methodology.

Dataset acquisition and preparation

A retrospective collection of 6,413 dental implant radiographs (periapical and panoramic) from 35 dental implant brands was compiled. The images were split into 70% for training, 15% for validation, and 15% for testing. The acquired images were cleaned, preprocessed, and annotated with ground truth using the Makesense.ai software. The dataset was resized to 640 × 640 pixels by adding padding.

The dataset included 35 implant systems. The number of samples per class varied and was not perfectly balanced because implant prevalence varies clinically. Weighted cross-entropy loss was employed to reduce bias toward majority classes. Data augmentation was also applied uniformly across classes.

To minimize potential bias toward majority classes, data augmentation was uniformly applied during training. Performance was also evaluated using precision, recall, and F1-score, thereby reducing the influence of dominant classes on overall performance. Nevertheless, class imbalance remains a potential limitation, particularly for implant systems represented by relatively few samples.

Detection and ROI extraction (YOLOv11)

In the first phase of the experimental pipeline, YOLOv11 served as the primary detection engine for automatically locating implants in full radiographic images. Figure 2 shows the YOLO region-of-interest (ROI) extraction pipeline. YOLOv11 was chosen for its more powerful architectural components, especially the addition of C3k2 blocks and C2PSA modules, which enhance feature extraction and greatly improve localization accuracy even in intricate medical backgrounds. YOLOv11 was trained for 100 epochs. When the model successfully detected the implant, it output the exact bounding box coordinates, allowing the image to be automatically cropped into localized Regions of Interest (ROI). These raw detected crops, without any further changes, became the “Baseline” dataset for subsequent classification benchmarking.

Foundation model-based segmentation (SAM2)

To thoroughly assess the effects of noise removal, including suppression of alveolar bone density, neighboring dental roots, and metallic artifacts, the Segment Anything Model 2 (SAM2) was used within the pipeline. Initial tests were conducted by directly applying SAM2 to entire radiographs in a zero-shot mode. However, this method did not yield clinically accurate masks. It was characterized mainly by mask leakage into adjacent anatomical structures or by an inability to distinguish the implant from the dense alveolar bone. Hence, a combined method was used, in which YOLOv11 serves as a localized intellectual trigger, providing the necessary spatial constraints for segmentation.

In this improved workflow, the high-accuracy bounding box coordinates produced by YOLOv11 were used as sparse prompts for the SAM2 predict function. These bounding box coordinates (x1, y1, x2, y2) were converted into a Numpy-encoded box input, producing the spatial anchor (located around the metallic implant). Because these were localized prompts, SAM2 achieved high accuracy, generating high-fidelity binary masks with much higher IoU (Intersection over Union) scores than when no prompts were used. In addition, the VAOD (Visible Artifacts Optimal Density) was calculated using the highest-scoring mask, based on the model’s internal confidence logits, to achieve the most accurate representation of the implant body.

Post-processing of the segmentation used a suppress-the-background technique (setting background pixels to zero while retaining the implant’s original pixel values). This was achieved by deriving an automated tight crop of the mask from the image’s maximum and minimum row and column indices. This produced an image with the background removed and tightly cropped to the implant, with a resolution of 512 × 512 pixels. This ensured that the structural elements of the implant were the only features available to the downstream DNN (Deep Neural Network) for classification, eliminating environmental noise from the classification dataset that could cause overfitting or misclassification in at least some downstream feature-extraction algorithms.

Generative super-resolution (ESRGAN)

To mitigate the spatial resolution limitations of clinical dental radiographs, a generative enhancement stage was introduced using the Real-ESRGAN framework. Initial experiments compared diverse image enhancement and super-resolution methods, including standard Bicubic interpolation, Contrast Limited Adaptive Histogram Equalization (CLAHE), and traditional Super-Resolution Convolutional Neural Networks (SRCNN). Although these techniques slightly improved contrast, they failed to restore the crisp edge definition necessary for thread-pitch analysis. Qualitative visual assessment indicated that reconstructing the implant microgeometry with Real-ESRGAN was the most clinically accurate and visually superior option. The method was highly effective at sharpening features without the halo artifacts typically produced by conventional sharpening filters.

The final pipeline uses the Residual-in-Residual Dense Block (RRDBNet) architecture. It contains 23 blocks, each with 64 features and 32 growth channels. This deep residual architecture enables the model to learn complex mapping functions from low-resolution (LR) to super-resolved (SR) space. Pre-trained weights from RealESRGAN_x4plus were used to achieve 4x scaling of the isolated implant areas. The input images were first converted to a 3-channel RGB stack and then processed with a RealESRGAN engine. To handle GPU memory limitations while preserving high-fidelity output across different regions of interest (ROI), a tiling strategy with a tile size of 256 and a padding of 10 was implemented. Figure 3 compares low-resolution images with the ESRGAN-applied super-resolution image.

The generator was configured to run half-precision (FP16) inference on CUDA-enabled hardware to accelerate the refinement process without compromising structural details. At this stage, accurately recovering the implants’ morphological fingerprint was critical, particularly through refinement of the platform-to-abutment interface and the thread lead angles. To validate the reliability of the generated attributes, the super-resolved images were quantitatively compared with their low-resolution counterparts. Finally, the enhanced data was used with the ConvNeXt and ResNet classification backbones to assess whether this visual enhancement improved diagnostic accuracy or, conversely, introduced architectural instabilities due to GAN-generated textural artifacts.

Experimental approaches for comparative analysis

To systematically assess how segmentation and generative enhancement affect the preprocessing paradox in relation to CNN performance, results from earlier stages were organized into four experimental configurations. In Configuration I (labeled Baseline), individual YOLOv11 crops containing the implants, along with the surrounding native anatomical background, were used. For Configuration II, the same implants were obtained using SAM2 to create masked versions, with the implants precisely segmented out against a black zero-pixel background, thereby removing all background anatomical details. Configurations III and IV examined the use of generative enhancement as an intermediate step before applying the more refined segmented background. For Configuration III (labeled Enhanced), Real-ESRGAN was applied to the raw YOLOv11 crops. In Configuration IV (labeled Fully Integrated), the Real-ESRGAN generative network was applied to the SAM2-segmented masks before feeding them into the classification CNNs. These four dataset configurations were used to train and test both the ConvNeXt and ResNet architectures.

Classification architectures and evaluation

The baseline: ResNet-50

ResNet-50 was chosen as the standard baseline architecture for this work because of its time-tested efficiency in medical image classification and its ability to work with deep feature hierarchies. The breakthrough of ResNet-50 is the Residual Block, which uses identity skip connections to address the vanishing gradient problem.

In mathematical terms, if the intended mapping is h (x), the residual block will learn the mapping f (x) = h (x)− x. The resultant output will be f (x) + x; thus, the network can pass the input directly to the layers. When it comes to dental implants, this architecture is especially good at preserving the original textural signature of the metal threads. A typical four-stage bottleneck setup was used, where each stage increases the channel depth and reduces the spatial dimensions via strided convolutions.

The modern standard: ConvNeXt-tiny

To explore how contemporary architectural designs influence generative artifacts, ConvNeXt-Tiny was chosen. While ResNet adheres to the traditional design conventions of the 2010s, ConvNeXt essentially rebuilds a typical Convolutional Neural Network (CNN) to capture the global perspective of a Vision Transformer (ViT) while preserving the inductive bias of convolutions. Macro Design: ConvNeXt replaces the classic ResNet stem with a “Patchify” layer (a 4 × 4 convolution with stride 4), enabling the network to capture broader spatial relations early in feature extraction. Extending Kernel Size: While ResNet uses 3 × 3 kernels, ConvNeXt uses 7 × 7 depthwise convolutions. This greatly expands the Effective Receptive Field (ERF), enabling the model to map the apical anatomy of an implant to its coronal platform with greater accuracy. Inverted Bottleneck and LN: The design adopts the Inverted Bottleneck style (a common feature in Transformers) and replaces Batch Normalization with Layer Normalization (LN). It was assumed that this change would improve stability when handling the artificially created texture produced by the ESRGAN upscaler.

Training protocol and optimization dynamics

Both models were fine-tuned with the AdamW optimizer, which separates weight decay from the gradient update, thereby enabling stable convergence with modern backbones. To tackle the challenging task of processing highly contrasting implant-segmented images (black background), our experiments were conducted using Automatic Mixed Precision (AMP). This approach lets the networks perform calculations in 16-bit floating-point (FP16) while maintaining a 32-bit (FP32) master copy of the weights. Hence, it doubles the processing speed and reduces memory usage on NVIDIA GPUs.

Evaluation methodology and tiered benchmarking

To quantify how closely the images after each enhancement step matched the original, the pipeline used objective image quality assessment metrics. Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) were calculated to quantify the quality of super-resolved images relative to the original YOLO-detected crops. PSNR captured pixel-level error and the amount of noise introduced during the generative process. In contrast, SSIM provided a score more aligned with human perception by assessing how well the image’s texture, luminance, contrast, and structural features were preserved. These measures provided additional evidence that Real-ESRGAN upscaling did not produce structural hallucinations that would mislead diagnostic interpretation of morphological features. Top-1 Accuracy and F1-score were used to evaluate the performance of the baseline, segmented, and enhanced levels, providing insight into how the model architectures were affected by the generative pre-processing steps.

The experimental evaluation was divided into three phases: automated detection, generative enhancement, and architectural robustness, to assess the effects of each change. The main aim of the research was to demonstrate Architecture Sensitivity by comparing raw radiographic crops with the complete SAM2 and ESRGAN pipeline. All experiments were conducted on an NVIDIA T4 Tesla GPU via Google Colab.

Localization and detection performance (YOLOv11)

The YOLOv11 object-detection model established a high-fidelity baseline for the entire pipeline, accurately localizing dental implants across diverse radiographic conditions, including varying bone densities and prosthetic superimpositions. The model achieved a Mean Average Precision (mAP@0.5) of 0.976 and an F1-score of 0.94. Figure 4 shows plots of the F1-score and map@0.5. The model’s tightly fitting bounding boxes enabled consistent Region of Interest (ROI) extraction, a key factor in ensuring that the subsequent segmentation and super-resolution stages were not affected by excessive anatomical noise. In Figure 5, the YOLOv11 detection results show precise bounding boxes.

Image Quality Assessment (IQA)

To quantify the effect of Real-ESRGAN upscaling, full-reference metrics were used to measure the difference between the original 512p crops and the super-resolved outputs.

PSNR: The pipeline delivered an average Peak Signal-to-Noise Ratio of 34.64 dB, indicating fairly controlled pixel-level variation throughout the generative process.

SSIM: The Structural Similarity Index was 0.9531, indicating that the major features of the implant (shape, number of threads, and platform width) were structurally preserved.

The enhancement step delivered a considerable Average Sharpness Improvement of 195.12. The Average Contrast Improvement of 1.36 and the Edge Strength Improvement of 0.87 further indicate that the ESRGAN model has effectively recovered high-frequency details, which are typically lost in standard clinical X-rays. Moreover, visual inspection showed that while traditional interpolation methods blurred the thread-led angles, the ESRGAN step restored edge sharpness, resulting in a more visually pleasing and suitable image for a potential clinician’s review. Figure 6 compares the raw ROI image, the segmented image, and the super-resolved image.

Comparative classification performance

The core of the research was benchmarking ResNet-50 against ConvNeXt-Tiny across four experimental levels. The outcomes, briefly presented in Table 1, reveal a significant difference in how classic and modern architectures handle generative preprocessing.

At level 1, ConvNeXt showed an inherent advantage over ResNet, likely due to its larger 7 × 7 kernels. However, the most important figure was at level 4, where the ResNet model experienced a 15% absolute drop in accuracy compared to its baseline. By contrast, ConvNeXt was robust, losing only 1% of its baseline accuracy. This indicates that ConvNeXt is much more stable against hallucinations and high-frequency textures, which are characteristic of GAN-based upscaling. Figure 7 shows the accuracy comparison of CNNs across all four levels. Figure 8 shows the normalized confusion matrix for ConvNeXt, demonstrating strong overall performance: most implant systems achieve high accuracy along the diagonal, indicating correct classification. Misclassifications were observed among implant systems with similar radiographic characteristics, whereas distinct implant designs, such as Astraev, Straumann Tissue Level, and Bicon, performed well. The results indicate that the model classified most implant systems, though challenges were noted with morphologically similar ones.

Statistical analysis of robustness

The difference in performance between the two models at the integrated level was statistically significant (p < 0.01). SAM2-mediated segmentation (level 2) led to a general decrease in accuracy, suggesting that, in addition to the bone-implant interface, a secondary diagnostic signal was also being used. The addition of ESRGAN only negatively impacted the model that lacked a modern hierarchical receptive field (ResNet). ConvNeXt effectively filtered out synthetic noise, enabling it to focus on global structural features.

According to the 95% confidence interval analysis, ConvNeXt-Tiny outperformed ResNet-50 across all four experimental levels. The narrow confidence intervals indicate stable performance with limited variability. Cohen’s kappa coefficients indicated substantial to almost good agreement between the predicted implant classes and the ground-truth annotations. The McNemar’s test also indicated that ConvNeXt-Tiny outperformed ResNet-50 across all experimental configurations. The statistical evaluation is summarized in Table 2, which reports classification performance, 95% confidence intervals, and Cohen’s κ values for both models across the four experimental tiers. Table 3 presents pairwise statistical comparisons using McNemar’s test.

Computational efficiency

The integrated pipeline required an average of 2 s per image for full processing on an NVIDIA Tesla T4 GPU. For 6,413 images, the high PSNR (34.64 dB) indicates that the pipeline is not only accurate but also consistent in output quality, making it suitable for high-throughput clinical deployment.

Ablation study

The proposed framework was evaluated through a structured ablation study with four experimental configurations. Rather than evaluating only the final integrated pipeline, each preprocessing component was introduced incrementally to quantify its individual contribution to implant classification performance. This design enables a systematic assessment of how ROI extraction, segmentation, and super-resolution affect the robustness of deep learning classifiers.

Tier 1: Baseline (YOLOv11 + CNN).

The baseline configuration used YOLOv11 to detect dental implants and automatically extract regions of interest (ROIs) from radiographic images. These ROIs were classified directly using ResNet-50 and ConvNeXt-Tiny, without additional preprocessing. ConvNeXt achieved 83% accuracy, outperforming ResNet (79%), indicating that larger receptive fields provide better discrimination among implant systems.

Tier 2: YOLOv11 + SAM2 Segmentation.

The second configuration used SAM2-based segmentation to isolate the implant by removing surrounding anatomical structures. This experiment was designed to assess whether eliminating background information would improve classifier performance. Contrary to expectations, both classifiers showed lower accuracy (ConvNeXt: 83 to 81%; ResNet: 79 to 76%).

Tier 3: YOLOv11 + ESRGAN Enhancement.

The third configuration evaluated the effect of ESRGAN-based super-resolution while preserving the original anatomical context. Although ESRGAN substantially improved image sharpness and achieved high PSNR and SSIM values, classification performance remained essentially unchanged (ConvNeXt: 83%, ResNet: 80%). This result suggests that improved perceptual image quality alone does not necessarily enhance the discriminative features learned by convolutional neural networks.

Tier 4: Fully Integrated Pipeline (YOLOv11 + SAM2 + ESRGAN).

The final configuration integrated ROI detection, segmentation, and super-resolution into a single preprocessing pipeline. This experiment evaluated the cumulative effect of all preprocessing stages. While ConvNeXt maintained robust performance with only a marginal decrease in accuracy (82%), ResNet saw a substantial reduction (64%). These findings demonstrate that modern convolutional architectures are considerably more robust to preprocessing-induced artifacts than traditional residual networks and provide quantitative evidence for the proposed preprocessing paradox.

Dental radiographs pose significant challenges due to anatomical overlay, noise, and the consistent geometry of contemporary implants. Many previous strategies have been unsuccessful in addressing these limitations in real-world clinical datasets, so the current study developed a rigorous, sequential preprocessing pipeline to address them. First, anatomical background noise was removed using YOLOv11 and SAM2. Second, lost resolution and detail were synthetically recovered through ESRGAN. Third, a comparison of ConvNeXt against ResNet50 was performed across 35 implant classes to determine whether the use of hyper-refined, artificially enhanced clinical images affects model performance or whether the creation of generative artifacts weakens architectural performance.

The results of the study indicate that the pipeline can generate “clinician-ready” images, as evidenced by an Average Sharpness Improvement of 195.12, a high-fidelity PSNR of 34.64 dB, and an SSIM of 0.9531. However, the subsequent classification results showed the human-supportive nature of visual enhancement, whereas the features extracted by machine learning were markedly different. This difference highlights a Preprocessing Paradox: eliminating anatomical noise and adding generative textures that benefit human visual clarity and architectural stability can destabilize long-established convolutional architectures, and visual fidelity is not a universal improvement in diagnostic accuracy.

The results of the current study reveal that image enhancement with ESRGAN does not necessarily improve CNN diagnostic accuracy. The assumption that better-looking images lead to better model performance may not always hold. Artificial enhancement of implant images may introduce non-biological information that does not add clinically meaningful information. The study reveals an important methodological insight: implant identification typically depends on the structural implant rather than on image clarity. ESRGAN definitely enhances visual quality, but not necessarily the distinctive features that CNNs require.

The reason is that generative up-scaling only adds visually appealing high-frequency textures to the edges, and these artificial pixels will appear as noise to traditional feature extractors. The accuracy of the ResNet-50 drops to only 64% on fully enhanced images. In contrast, ConvNeXt is less sensitive to visually appealing artifacts and therefore offers better prospects for human review of enhanced images.

The “contextual signal” hypothesis (the SAM2 effect)

A critical finding was a drop in performance after SAM2-mediated segmentation (level 2), with accuracy decreasing for both models (ConvNeXt: 83 to 81%; ResNet: 79 to 76%). This finding supports the “Contextual signal” explanation, which proposes that deep learning models derive diagnostic information from the bone-implant interface, the distinct radio-opacity patterns of osseointegration, and the implant’s relative position to the cortical plate (Sharma et al., 2020). In clinical settings, the rationale for surgical placement (e.g., subcrestal vs. epicrestal) serves as a secondary brand identifier. By using SAM2 to isolate the metallic silhouette, the anatomical context that neural networks use to differentiate between brands with similar macro-geometries is unintentionally removed (Lashaki et al., 2025). However, it is also possible that imperfect segmentation masks distort implant thread morphology, thereby degrading the discriminative features used by the classifier. Therefore, the decline in performance may be due to segmentation-induced distortions along with the loss of contextual information. Segmentation accuracy was not evaluated separately in the study; the contribution of these two factors warrants further investigation.

The preprocessing paradox (the ESRGAN effect)

The Real-ESRGAN implementation introduced a preprocessing paradox. The generative model alone achieved a structural similarity of 0.9531, indicating it almost perfectly preserved structure while adding high-frequency synthetic textures. Visually, these textures appeared crisper to humans, with more distinct thread details. However, the automated classifier treated them as non-adversarial noise. This difference implies that although GAN-based upscaling performs well in human visual inspection, it can generate stochastic hallucinations that disrupt the typical weight distributions of traditional convolutional layers (Rombach et al., 2022).

Architectural robustness: ConvNeXt vs. ResNet

The most notable finding of the study was the dramatic drop in ResNet-50 accuracy to 64% in the integrated tier, while ConvNeXt remained steady at 82%. The following key architectural differences can explain this disparity: Receptive Fields: ResNet-50 uses small 3 × 3 kernels that focus on local textures. When ESRGAN alters these local textures, ResNet is fooled by the resulting artifact features (He et al., 2016). By contrast, ConvNeXt uses 7 × 7 kernels and a patchify stem, resulting in a larger Effective Receptive Field (ERF) that captures global macro-geometry more than local synthetic noise Liu et al., 2022. Normalization Layers: ResNet relies on Batch Normalization, which can struggle with the sparse, black-background data generated by SAM2 (Adnan et al., 2025). ConvNeXt employs Layer Normalization, which improves stability when processing isolated, segmented ROIs. However, we did not perform Grad-CAM heatmaps, which were not used to illustrate the decline in ResNet-50 accuracy. Future studies should employ these techniques to better understand the features that influence ResNet-50 predictions after ESRGAN enhancement and segmentation.

Comparison with existing literature

The diagnostic performance of the YOLOv11 + ConvNeXt pipeline marks a major leap over previous benchmarks, especially given the number of classes and the architecture’s strength. Early studies were mostly focused on binary detection or on very few idealized systems. There are about 300 implant classes worldwide, but to properly train a Convolutional Neural Network, a large, balanced dataset is needed, which is rarely available for rare or older brands. The current study, focusing on the 35 most clinically relevant systems, effectively reflects the real market to a large extent. Not only does this method solve the most common diagnostic problems doctors face, but it is also a highly flexible system that can easily add new classes as more radiographic data are collected.

Figure 9 illustrates how the proposed multi-stage pipeline works in practice to achieve automated dental implant identification. As shown in Figure 9, existing works (Park et al., 2023; Kong et al., 2025; Leblebicioglu Kurtulus et al., 2024; Park et al., 2023; Yuce et al., 2025; Kong, 2023; Hassan et al., 2024; Pachaiappan et al., 2026) in the dental implant identification literature indicate that, as dataset complexity increases, classification accuracy decreases in an inverse relationship. Models can classify with very high accuracy, close to 100%, within 3–9 classes, and drop to a normal value in the low 80% when working with the 25 classes of dental implants. The present study bridges this gap by systematically evaluating generative preprocessing trade-offs on a massive multi-class dataset. By maintaining a robust 83% baseline accuracy across an unprecedented 35 distinct commercial systems, the proposed YOLOv11 + ConvNeXt pipeline significantly advances the state of the art for clinical-scale identification.

As the number of classes increases, the inter-class distance in the feature space decreases, and different brands start to appear similar at low resolutions. The proposed pipeline outperforms implant differentiation, producing macro-geometries highly similar to implants while maintaining 83% accuracy (Tier 1) and 82% accuracy (Tier 4) across 35 systems.

Moreover, the current study uncovers an important technical issue that is rarely discussed in the literature: the Preprocessing Paradox. While Lee et al. (2020) and similar studies have only evaluated the effectiveness of YOLO-based detection, this study shows that combining generative upscaling (ESRGAN) with ResNet-50 can lead to a 15% drop in performance. In contrast, using modern backbones such as ConvNeXt can help maintain stable performance. This comparative finding is crucial for the evolution of clinical tools, as it demonstrates that architectural robustness is as important as image quality in automated dental diagnostics.

Although the proposed framework demonstrated promising performance on a retrospective dataset of 6,413 radiographs from 35 dental implant systems, all experiments were conducted using data from multiple institutions. Consequently, variations in imaging equipment, acquisition protocols, patient demographics, and implant prevalence across clinical centers were not evaluated. At present, publicly available annotated multiclass dental implant datasets suitable for external validation are limited, preventing independent validation of the proposed methodology. Therefore, the generalizability of the proposed framework should be interpreted with caution. Future work will focus on validating the model using multicenter datasets from different institutions, incorporating domain adaptation techniques, and evaluating robustness across diverse clinical environments to facilitate clinical translation.

Clinical implications, limitations, and future directions

For a clinician or forensic odontologist, the Sharpness Improvement is highly useful for manual identification when records are lost. However, the current experiments indicate that clinical AI developers should avoid “over-preprocessing” in automated systems; the baseline YOLOv11-ConvNeXt yields the most trustworthy results.

A primary limitation is the study’s 2D nature; periapical radiographs are prone to projection geometry, angulation, magnification, and anatomical superimposition, which distort the thread-pitch metrics that ESRGAN aims to improve. However, ESRGAN and SAM2 cannot compensate for missing information. Consequently, variations in radiographic acquisition may have affected classification performance. Future studies should evaluate the proposed framework using multi-center datasets with greater imaging variability and investigate its applicability to three-dimensional imaging modalities, such as cone-beam computed tomography (CBCT), where appropriate. Future research should investigate Latent Diffusion Models, which may provide more consistent textural restoration than GANs, and extend this pipeline to 3D-CBCT volumetric data to eliminate geometric distortion.

The study outlined a multi-stage deep learning pipeline for identifying dental implants across 35 different systems. The combination of SAM2 and Real-ESRGAN not only enhanced the visual quality but also the structural fidelity of the radiographs, achieving a PSNR of 34.64 dB. However, the study exposed a vital Preprocessing Paradox in automated diagnostics. Artificial enhancement of dental images does not necessarily improve the accuracy of CNNs for implant identification. The study suggests that current transformer-inspired architectures, such as ConvNeXt, are inherently robust and maintain high diagnostic accuracy (82%) even in the presence of generative artifacts. In contrast, typical residual networks (ResNet) experience a total breakdown (64%). Therefore, this underscores the importance of selecting modern convolutional backbones for the next generation of dental AI systems.