Interpretable Deep Learning Models for Breast Cancer Diagnosis Using Medical Imaging

1 Introduction

• Limited annotated data, substantial computational needs, and model opacity remain challenges for deep-learning breast cancer diagnosis.Understanding model decisions matters in healthcare because erroneous predictions can have severe repercussions.Source: §1 · p. 2
• Transfer learning uses pretrained models to reduce training demands on specialized medical datasets, while XAI methods visualize model behavior.The paper notes that these methods may enhance clinician trust and streamline regulatory approval.Source: §1 · p. 2
• The paper proposes an interpretable framework combining transfer-learning models with explainable AI for breast cancer classification.It assesses ten architectures and incorporates augmentation, normalization, patch-based preprocessing, and methods for local and global interpretability.Source: §1 · p. 2
• The paper is organized to cover prior work, datasets and preparation, methods, experiments, results, and conclusions with limitations and future directions.Source: §1 · p. 2

2 Literature Review

• Silva-Aravena et al. [2023][14] reported 85% accuracy using XGBoost on 400 Indonesian participants, with SHAP identifying risk factors.The authors noted that the limited, self-reported dataset constrained generalisability.Source: §2 · pp. 2–3
• Nasir et al. [2025][15] reported 98.2% accuracy for CNN+LSTM on 569 tabular cases, while noting limited dataset size and clinical-imaging generalizability.Grad-CAM and SHAP were used for interpretability.Source: §2 · p. 3
• Patho-Net reported 98.9% accuracy on BreakHis histopathology images, but was limited to binary classification and resource-intensive preprocessing.Grad-CAM and XRAI augmented interpretability.Source: §2 · p. 3
• Other studies reported 91.02% accuracy for EfficientNetV2S on BUSI, variable DenseNet169 accuracy across datasets, and 85.77% accuracy for ultrasound stacking.These studies also noted constraints including limited or imbalanced datasets, insufficient validation, benign-case errors, or absent XAI.Source: §2 · p. 3
• Sharmi et al. [2025] combined Attention U-Net, K-Means, EfficientNet-B7, and LIME for thermal-image diagnosis, reporting 91.7% validation accuracy.Class imbalance, demographic bias, and insufficient cross-validation constrained generalizability.Source: §2 · p. 3

3 Research Gap

• Breast-cancer deep-learning research often prioritizes accuracy over interpretability, limiting medical applicability.Source: §3 · p. 3
• XAI integration is described as superficial and inconsistent across architectures, while reliance on a single dataset limits generalizability.Source: §3 · pp. 3–4
• Few benchmarks compare CNN and transformer models, and augmentation and preprocessing effects on interpretability remain little examined.The study responds with an assessment of transfer-learning models using several XAI methods.Source: §3 · p. 4

4.1 Dataset

• BreakHis contains 7,909 images from 82 patients at four magnifications, split by patient into training, validation, and test sets.The 70/20/10 split prevents patient overlap; weighted sampling addresses class imbalance, though intra-patient and inter-magnification similarities may remain.Source: §4.1 · p. 4

4.2 Data Augmentation

• Augmentation uses rotations, flips, shearing, zooming, and contrast adjustment to represent imaging variability while aiming to preserve diagnostic integrity.Bilinear interpolation is used during scaling and translation.Source: §4.2 · p. 4

4.3 Transfer Learning

• Pretrained models are adapted by replacing their final classification layers and fine-tuning selected layers on the target dataset.The paper evaluates ResNet, VGG, DenseNet, EfficientNet, Xception, MobileNet, and vision-transformer architectures.Source: §4.3 · p. 4

4.3.1 DenseNet121

• DenseNet121 uses dense connectivity and a custom 1024→512→2 classifier head for benign-versus-malignant logits.It is fine-tuned with label-smoothed cross-entropy, AdamW, differential learning rates, cosine warm restarts, RandAugment, and gradual unfreezing.Source: §4.3.1 · pp. 4–5

4.3.2 EfficientNet

• EfficientNet models use compound scaling, MBConv blocks, and squeeze-and-excitation modules, with custom binary-classification heads.The study fine-tunes ImageNet-pretrained B0, V2S, and B7 variants, progressively unfreezing the backbone.Source: §4.3.2 · p. 5

4.3.3 ConvNeXt

• ConvNeXt-Tiny/Base combines CNN locality and translation equivariance with transformer-inspired design elements for binary classification.Its modified head uses LayerNorm, GELU, dropout, and fully connected layers; training includes gradual unfreezing and test-time augmentation.Source: §4.3.3 · p. 5

4.3.4–4.3.6 and 4.3.10

• ResNet-50 and InceptionV3 use ImageNet-pretrained backbones with custom binary heads; InceptionV3 also uses multi-scale filters and a modified auxiliary classifier.VGG-19 replaces its original head with a large multilayer classifier, while Xception uses depthwise separable convolutions and residual connections.Source: §4.3.4 · p. 5, §4.3.5 · pp. 5–6, §4.3.6 · p. 6

4.3.7–4.3.9

• MobileNetV2 uses depthwise separable convolutions and inverted residual bottlenecks; ViT processes 16×16 image patches as tokens, while Swin uses shifted-window attention and patch merging.Their custom heads produce benign-versus-malignant classifications; MobileNetV2 has about 3.5 million parameters, ViT over 86 million, and Swin about 28.3 million.Source: §4.3.7 · p. 6, §4.3.8 · p. 6, §4.3.9 · p. 6

4.4.1 Grad-CAM (Gradient-weighted Class Activation Mapping)

• Grad-CAM uses class-score gradients from the final convolutional layer to weight feature maps and produce heatmaps highlighting image regions important to classification.The heatmap is overlaid on the image; the method is post hoc and requires no architectural changes.Source: §4.4.1 · p. 7

4.4.2 LIME (Local Interpretable Model-Agnostic Explanations)

• LIME perturbs image superpixels and fits a local linear surrogate to show which segments influence an individual prediction.It offers instance-specific explanations that can complement global saliency maps by showing local decision limits.Source: §4.4.2 · p. 7

4.4.3 SHAP (SHapley Additive Explanations)

• SHAP estimates Shapley values to allocate prediction influence among image superpixels or patches and visualize influential regions.The section contrasts its globally consistent explanations with gradient-based techniques.Source: §4.4.3 · p. 7

4.4.4 Integrated Gradients (IG)

• Integrated Gradients averages pixel-level gradients along a path from a baseline image to the input to attribute the prediction.The method is applied to CNN and transformer architectures to examine tissue patterns relevant to benign–malignant classification.Source: §4.4.4 · p. 7

4.4.5 Layer-wise Relevance Propagation (LRP)

• LRP redistributes prediction scores backward through the network to identify input pixels responsible for the decision.It is applied to CNNs including ResNet, DenseNet, and Xception to visualize histological features.Source: §4.4.5 · pp. 7–8

6.1 Datasets

• BreakHis contains 7,909 images from 82 patients, with benign and malignant classes, eight subtypes, and four magnifications from 40X to 400X.The dataset has class imbalance and substantial within-class variability.Source: §6.1 · p. 8

6.2 Experimental Results

• 99.41% test accuracy and 0.9997 AUC-ROC made DenseNet-121 the top model; it also achieved 100% across measures at 40X, 100X, and 200X.Its test precision was 99.86%, recall 99.45%, specificity 99.17%, and F1-score 99.65%.Source: Figure 37 · p. 26, Figure 38 · p. 27, §6.2 · p. 27
• Xception, Inception V3, and EfficientNet were competitive CNNs, while ViT and Swin trailed CNNs and ConvNeXt had 2.42% specificity and 77.14% test accuracy.Xception achieved 85.84% test accuracy; transformer models had approximately 77–79% test accuracy, and ConvNeXt showed a bias toward malignant predictions.Source: §6.2 · p. 27
• Transformer models retained recall above 83% across magnifications but had lower precision and specificity than the leading CNNs.The authors identify further improvement for domain-specific heterogeneity as necessary for vision transformers.Source: §6.2 · p. 27

7 Explainability

• DenseNet121 explanations localized nuclei and glandular structures, with Integrated Gradients, LIME, and LRP highlighting diagnostically relevant tissue evidence.Grad-CAM localized discriminative regions progressively in deeper layers; Integrated Gradients highlighted nuclei-dense areas and boundaries across baselines, while LRP separated positive epithelial and nuclear contributions from negative background relevance.Source: §7 · pp. 27–28, Figure 39 · p. 28
• EfficientNet, InceptionV3, and MobileNetV2 explanations also emphasized nuclei-rich, glandular, epithelial, or nuclear features through complementary XAI methods.EfficientNet’s deeper Grad-CAM layers focused increasingly on nuclei-rich and glandular structures; InceptionV3 shifted from texture toward nuclei-dense areas in deeper modules; MobileNetV2 shifted from textures toward nuclei-packed regions. Integrated Gradients, LIME, and LRP or saliency methods supplied complementary attributions.Source: §7 · pp. 28–30, Figure 40 · p. 29, Figure 41 · p. 30

7.1 Statistical Significance, Robustness, and Generalization Analysis

• 84.6% to 86.2% accuracy and 0.933 to 0.940 AUC characterized the second tier of EfficientNet, Xception, InceptionV3, and MobileNetV2.EfficientNet had the lowest magnification variability in this group (σ = ±0.75%) and Friedman rank 3.62; Xception had 91.2% sensitivity, while InceptionV3 and MobileNetV2 showed similar pairwise performance.Source: §7.1 · pp. 31–32
• ResNet50 and VGG19 had lower Friedman rankings, with magnification sensitivities of σ = ±2.36% and ±1.99%, respectively.Their agreement coefficients were 0.625 and 0.573, and their MCC values were 0.645 and 0.599, respectively.Source: §7.1 · p. 32
• 99.4% average test accuracy and 0.999 AUC-ROC; DenseNet121 ranked first and statistically outperformed all other models (p < 0.001).Its magnification sensitivity was σ = ±1.08%.Source: §7.1 · p. 31, Figure 43 · p. 32, Figure 44 · p. 32
• 79.5% and 77.6% accuracy, respectively, and AUC values below 0.86 characterized Swin Transformer and ViT; CNNs were more accurate than transformers (p = 0.0032).CNNs also had higher AUC (p = 0.0047); the transformers showed the largest magnification sensitivities, σ = ±2.99% and ±3.18%.Source: §7.1 · p. 32
• 77.1% accuracy, 0.653 AUC, and -0.180 MCC made ConvNeXt the weakest architecture by these measures.Its low κ (0.477) and high AUC dispersion across magnifications also indicated poor discrimination and instability.Source: §7.1 · p. 32
• Statistical validation combined classification metrics, significance tests, agreement measures, confidence intervals, and robustness analyses to corroborate the reported performance hierarchy.The supplied figure captions identify comparisons across the patient-stratified test set, architectures, and magnification levels.Source: Figure 43 · p. 32, Figure 44 · p. 32, §7.1 · p. 33

8 Limitations

• A single histopathological dataset and binary classification constrain generalization to other imaging modalities, subtype classification, grading, and staging.The authors also caution that near-perfect results may reflect dataset-specific correlations or some validation-set learning; XAI methods were mostly qualitative, and independent multi-institutional testing plus quantitative interpretability assessment is needed before clinical use.Source: §8 · p. 33

9 Conclusion & Future Work

• DenseNet121 achieved the highest accuracy (99.41%), AUC-ROC (0.9997), Cohen’s κ (0.986), and MCC (0.984), significantly outperforming all compared models (p < 0.001).It showed magnification sensitivity of σ = ±1.08%; Grad-CAM and LIME identified nuclei-dense and glandular regions for DenseNet121, EfficientNet, and Swin Transformer.Source: §9 · p. 33
• The authors identify transfer learning combined with explainable AI as a potential basis for accurate, transparent, clinically relevant breast cancer decision-support systems.They call for independent multi-institutional and multimodal validation, extension to subtyping and grading, and quantitative and clinical evaluation of interpretability.Source: §9 · p. 34

Author contributions statement

• Author 1 analyzed results and reviewed the draft; Authors 2 and 4 contributed to methodology, drafting, and data analysis; Author 3 conducted experiments and generated results.Authors 5 and 6 handled project administration and results evaluation, and all authors reviewed, edited, and approved the final manuscript.Source: Author contributions statement · p. 34

• The authors conducted a secondary analysis of publicly available, de-identified BreakHis histopathology images.Source: Ethics Statement · p. 34
• The study involved no direct human participation, patient recruitment, or intervention by the authors.Source: Ethics Statement · p. 34
• Because the dataset was collected under relevant clinical and ethical guidelines and its images were anonymized, additional ethical approval and informed consent were not required.Source: Ethics Statement · p. 34