PaperlayerView exact revision on arXiv

Source-linked AI summary

Architecture-aware Robustness Evaluation of Explainable Deep Learning for Breast Cancer Diagnosis

Balenthira Thanusanth, Selvarajah Thuseethan, Roshan G. Ragel, Bimali S. Weerakoon, Ayesh Jayasinghe, Ananthamoorthy Krishnamoorthy

arXiv:2609.31692v1cs.CV

TL;DR

Breast cancer XAI selection remains ad hoc, with limited systematic comparison across model architectures. This study evaluates methods across architectures and finds explanation quality depends on the model–method interaction, rather than one technique consistently prevailing.

Abstract

from arXiv · show

Explainable Artificial Intelligence (XAI) has become essential in medical image analysis to ensure transparency of deep learning (DL)-based diagnostic systems. However, selecting appropriate XAI techniques for breast cancer recognition remains largely ad hoc, with limited systematic evaluation across different DL architectures. This study presents a systematic architecture-aware evaluation protocol to assess the effectiveness of nine widely used XAI techniques across four categories of DL models: very deep, lightweight, transformer-based and hybrid neural networks. The evaluation is conducted on a breast ultrasound dataset comprising 780 images using clinically aligned spatial metrics, including Pointing Game, Intersection over Union and Mean Coverage, to quantify agreement between generated explanations and expert-annotated lesion regions. Results indicate that explanation quality is primarily influenced by the interaction between model architecture and XAI method, rather than any single technique consistently outperforming others. Hybrid architectures produce more spatially coherent explanations, while lightweight and transformer-based models exhibit greater variability across methods. The findings show that no single technique generalises across architectures and evaluation criteria, emphasising the need for joint selection of DL models and XAI techniques. Explainability depends on both model design and explanation strategy and should not be considered independently. This work provides a structured evaluation protocol and practical guidance for selecting XAI techniques in breast cancer diagnosis, supporting more transparent clinical decision-support systems. \textcolor{blue}{Code is publicly available at https://github.com/Nishan-Charlie/Explainable-AI}

I. INTRODUCTION

The section motivates systematic, architecture-aware XAI evaluation for breast cancer recognition, where method selection and evaluation are often inconsistent. It introduces a comparative framework intended to guide selection across models and methods.

I. INTRODUCTION

  • XAI selection is often unsystematic, with limited comparative evaluation across architectures and no clear guidance for matching methods to models.The section notes that methods may be applied without justification or standardized evaluation, potentially leaving explanations clinically misaligned.Source: §I · p. 1, A. Clinical Deployment Scenario and… · pp. 1–2

A. Clinical Deployment Scenario and Limitations

  • Clinical adoption calls for explanations that are understandable, clinically relevant, truthful, informative, and computationally efficient, although all five criteria are rarely achievable.In automated evaluation without expert involvement, explanation quality is assessed using quantitative metrics.Source: A. Clinical Deployment Scenario and… · p. 1
  • The framework compares XAI techniques across very deep, lightweight, transformer-based, and hybrid networks to assess their suitability for breast cancer recognition.The paper presents this as a response to architecture-independent method selection and the lack of standardized comparative evaluation.Source: A. Clinical Deployment Scenario and… · p. 2
  • The study evaluates nine XAI techniques across nine DL models spanning four architecture categories, using a breast ultrasound dataset and three quantitative metrics.The dataset comprises 780 images across three classes; the metrics assess explanation quality.Source: A. Clinical Deployment Scenario and… · p. 2
  • The work aims to provide empirical guidance for selecting an XAI technique for a given DL architecture in breast cancer recognition.The stated intended application is development of breast cancer decision-making systems in clinical settings.Source: A. Clinical Deployment Scenario and… · p. 2

II. RELATED WORK

The related work reviews established XAI methods and their use in breast cancer imaging, alongside deep-learning architecture families. It identifies limited systematic, multi-metric comparisons of XAI methods across architectures as an open gap.

A. XAI Benchmarks

  • Image-classification explainers include activation and gradient heatmaps, perturbation methods such as SHAP and LIME, occlusion, and sensitivity analysis.The review also describes Integrated Gradients and Guided Backpropagation as gradient-based approaches.Source: A. XAI Benchmarks · p. 2
  • Breast cancer studies use Grad-CAM variants, SHAP, and LIME for interpretation, with explanations sometimes compared against expert lesion masks using localization and overlap metrics.The review notes that combining explainers has also been proposed to produce richer explanations.Source: A. XAI Benchmarks · p. 2

B. Deep Breast Cancer Recognition

  • Very deep CNNs support hierarchical feature extraction but can make prediction features harder to trace, while lightweight models trade representational capacity for reduced complexity.The review associates reduced representation in lightweight models with potentially less detailed or discriminative explanations.Source: B. Deep Breast Cancer Recognition · p. 3
  • Transformers use self-attention to capture global context, but attention weights do not necessarily represent true feature importance and may yield inconsistent interpretations.The review describes the reliability of attention as an explanatory mechanism as contested.Source: B. Deep Breast Cancer Recognition · p. 3
  • Existing studies rarely compare benchmark XAI methods across diverse breast cancer architectures under a unified setting or evaluate them with multiple metrics.Prior work is often limited to a small set of methods, qualitative inspection, or single-model analysis.Source: B. Deep Breast Cancer Recognition · p. 3, A. XAI Benchmarks · p. 3
  • The literature identifies a need to compare XAI methods across very deep, lightweight, transformer-based, and hybrid networks using multiple assessment metrics.The section notes that current evidence across these architectures under a unified experimental setting remains limited.Source: B. Deep Breast Cancer Recognition · p. 3

III. PROPOSED METHOD

The method evaluates whether XAI explanation masks align with expert-annotated clinically relevant lesion regions across breast ultrasound images. It compares explanation methods across model architectures using spatial metrics.

A. Problem Formulation

  • Each trained model maps an ultrasound image to a diagnostic prediction, and each XAI method generates a saliency map for that model and image.Pixel values in the saliency map represent the importance of image regions for the prediction; the maps are normalised and binarised for comparison with the masks.Source: A. Problem Formulation · pp. 3–4
  • Explanation masks are assessed for alignment with expert-annotated lesion regions, treated as clinically relevant areas.The formulation describes this alignment as reflecting explanation faithfulness and clinical utility.Source: A. Problem Formulation · pp. 3–4
  • IoU, Pointing Game, and Mean Coverage assess explanation quality through spatial agreement, localisation accuracy, and lesion-region completeness.Pointing Game checks whether the most salient point falls inside the annotated region; Mean Coverage measures how much of that region the explanation captures.Source: A. Problem Formulation · p. 4
  • The evaluation seeks the most suitable XAI method for each architecture category by comparing methods under models in the architecture set.The framework conducts similar analyses for each XAI technique to enable comparative ranking.Source: A. Problem Formulation · p. 4, Figure 2 · p. 4

B. Benchmark XAI Techniques and DL Models

The study evaluates nine XAI techniques and nine deep-learning models spanning four architectural paradigms. The methods cover diverse attribution approaches, while the models represent varying computational and representational designs.

B. Benchmark XAI Techniques and DL Models

  • Nine XAI techniques spanning activation-, gradient-, perturbation-, surrogate-, and game-theoretic approaches are evaluated across four architectural paradigms.The model paradigms are very deep, lightweight, transformer-based, and hybrid neural networks.Source: B. Benchmark XAI Techniques and DL… · pp. 4–5
  • The benchmark includes CAM, Grad-CAM, Integrated Gradients, SHAP, LIME, Guided Backpropagation, Saliency Maps, Occlusion, and Sensitivity.The methods produce explanations through feature activations, gradients, perturbations, local surrogate approximation, or game-theoretic attribution.Source: B. Benchmark XAI Techniques and DL… · pp. 4–5
  • Very deep models include ResNet50 and DenseNet, while lightweight models include MobileNetV2 and ShuffleNet.The lightweight designs reduce computational cost for resource-constrained or real-time settings; transformer models use global or shifted-window self-attention, and hybrids combine convolution with attention.Source: B. Benchmark XAI Techniques and DL… · p. 5
  • Hybrid models combine convolutional operations with transformer attention to represent local patterns and global context.MobileViT, ConvNeXt, and CoAtNet are described as hybrid architectures in the model set.Source: B. Benchmark XAI Techniques and DL… · p. 5

C. Dataset and Evaluation Framework

The evaluation uses breast ultrasound images paired with lesion masks and compares explanations across trained model–XAI combinations. Metrics are averaged over test samples and used to rank methods for individual models and architecture categories.

C. Dataset and Evaluation Framework

  • The Breast Ultrasound Dataset contains 780 images across benign, malignant, and normal classes; this study uses benign and malignant samples.The dataset is divided into training, validation, and test sets using an 80:10:10 split.Source: C. Dataset and Evaluation Framework · pp. 5–6
  • Expert-reference lesion masks were created by freehand segmentation in MATLAB and paired with ultrasound images.The dataset description identifies the masks as references for evaluating XAI techniques.Source: C. Dataset and Evaluation Framework · pp. 5–6
  • All trained-model and XAI-method combinations are evaluated on test samples using IoU, Pointing Game, and Mean Coverage.Metric scores are averaged across samples, then used to rank methods per model and compare results aggregated by architecture category.Source: C. Dataset and Evaluation Framework · p. 6

IV. EXPERIMENTS

The section introduces the experimental setup, reports results from experiments assessing the proposed XAI evaluation framework, and provides discussion and recommendations.

IV. EXPERIMENTS

  • Experiments assess the proposed XAI evaluation framework, with results followed by discussion and recommendations.Source: §IV · p. 6

A. Experimental Setup

Experiments used PyTorch on a two-GPU server, with ImageNet-pretrained models trained using shared optimization and regularization procedures.

A. Experimental Setup

  • Experiments were implemented in PyTorch on a high-performance server with two NVIDIA RTX A6000 GPUs, each with 49 GB of memory.The environment used NVIDIA driver 550.144.03 and CUDA 12.4.Source: A. Experimental Setup · p. 6
  • Models were ImageNet-pretrained and trained with AdamW at a learning rate of 0.0001, with weight decay between 0.0003 and 0.0005 and batch size 16.Most models trained up to 300 epochs; CoAtNet-0 and MobileViT-S trained for 100 and 35 epochs, respectively.Source: A. Experimental Setup · p. 6
  • Training used dropout, label smoothing, cosine annealing for CoAtNet-0, and early stopping, with the best checkpoints selected on validation metrics.Dropout ranged from 0.2 to 0.4, label smoothing from 0.02 to 0.05, and early-stopping patience from 8 to 20 epochs.Source: A. Experimental Setup · p. 6

B. Results

Hybrid ConvNeXt led classification performance, while explanation usefulness also depended on whether the underlying prediction was correct.

1) DL Model Performance:

  • Hybrid models showed the most promising classification performance, while MobileViT recorded the lowest performance among the neural networks.Very deep models performed strongly, lightweight models were competitive, and transformer-based models showed similar results without a clear advantage.Source: 1) DL Model Performance · p. 6
  • 0.9494 accuracy and 0.9494 F1-score: ConvNeXt achieved the highest values among the evaluated models.ResNet50 reached 0.9367 accuracy and 0.9371 F1-score; Table I compares models by accuracy, precision, recall, and F1-score.Source: 1) DL Model Performance · p. 6, Table I · p. 7
  • ConvNeXt correctly classified both representative samples, whereas MobileViT misclassified the same benign and malignant cases.The MobileViT explanations were conditioned on incorrect predictions and could reflect miscalibrated decisions rather than pathology-specific evidence.Source: 1) DL Model Performance · pp. 6–7, Figure 4 · p. 7

2) Explainability Analysis:

XAI performance varied with architecture and evaluation metric, with hybrid models showing the strongest overall attribution patterns. No single explanation method consistently led across architectures and criteria.

2) Explainability Analysis:

  • For very deep networks, SHAP led PG, IoU, and MC; for lightweight models, SHAP led IoU and MC, while other methods led PG.Lightweight-model rankings were therefore fragmented across metrics, with no method consistently dominating.Source: 2) Explainability Analysis · p. 7
  • Transformer results were metric-dependent, while CoAtNet had the highest overall PG and IoU and Saliency and SHAP were strong performers.On ViT, SHAP was competitive in PG and IoU, whereas Occlusion and LIME achieved higher MC; ConvNeXt and MobileViT also showed competitive results with Integrated Gradients and Saliency.Source: 2) Explainability Analysis · p. 7
  • PG showed a wider spread across models than IoU and MC, whose scores were more compressed at higher performance levels.Figure 5 consolidates the best-performing methods across models and metrics, with colored backgrounds indicating architecture categories.Source: Figure 5 · p. 7, 2) Explainability Analysis · p. 7
  • SHAP and IG were more robust across most architectures, while CAM-based methods and LIME and GBP were weaker and less consistent, particularly on transformers.The overall pattern supports evaluating model–XAI compatibility rather than treating any method as intrinsically superior.Source: 2) Explainability Analysis · p. 8
  • Hybrid networks had the most favourable attribution scores, while lightweight and transformer-based models showed greater variability across methods.The heatmap presents average PG, IoU, and MC scores across models and methods; results indicate interaction-dependent rather than uniform performance.Source: 2) Explainability Analysis · pp. 7–8, Figure 6 · p. 8
  • For benign and malignant examples, hybrid models produced coherent lesion-aligned explanations, while some lightweight and transformer models showed fragmented or irrelevant attributions.These qualitative patterns were consistent with the quantitative findings and were illustrated across nine models for two representative samples.Source: 2) Explainability Analysis · pp. 8–9, Figure 7 · p. 9

C. Discussion and Recommendations

Explanation quality reflects the interaction between architecture and attribution mechanism, with learned feature representations constraining spatial coherence and clinical relevance. The discussion recommends architecture-aware method selection, multi-metric assessment, and checking explanation reliability alongside model correctness.

C. Discussion and Recommendations

  • Architecture–attribution interaction shapes explanation quality, as feature representations constrain spatial coherence and clinical relevance.Models capturing local texture and global context tend to produce more coherent explanations; less structured hierarchies yield less stable explanations sensitive to the XAI method.Source: C. Discussion and Recommendations · p. 9
  • Different XAI methods prioritize peak attribution, region coverage, or smoother spatial distributions, so a single metric may bias assessments.The discussion calls for multiple metrics to capture complementary dimensions and avoid overestimating visually appealing but incomplete explanations.Source: C. Discussion and Recommendations · p. 9
  • Practical deployment should select XAI methods by architecture and evaluate explanation reliability alongside model correctness.The discussion favors hybrid architectures for clinical explainability and identifies joint optimization of predictive accuracy and explanation faithfulness as future work.Source: C. Discussion and Recommendations · p. 9

D. Limitations

The conclusion finds that explanation quality depends on the architecture–XAI interaction, while the evaluation is bounded by one breast ultrasound dataset and quantitative spatial metrics. The authors identify expert assessment as future work.

D. Limitations

  • Evaluation on one limited breast ultrasound dataset from one clinical source constrains generalisability across modalities, protocols, and patient populations.The authors note that explanation behavior is sensitive to data distribution and may change under domain shifts.Source: D. Limitations · p. 10
  • Quantitative spatial metrics do not capture clinical interpretability, diagnostic relevance, or trustworthiness; the authors call for structured expert assessment.The conclusion identifies systematic medical-expert evaluation of clinical suitability and trustworthiness as future work.Source: D. Limitations · p. 10, §V · p. 10

V. CONCLUSION

  • Hybrid architectures produce more stable, clinically meaningful explanations, but no XAI method generalizes across all architectures and evaluation criteria.The study concludes that model design and explanation strategy must be considered jointly.Source: §V · p. 10

Related papers

Loading 2609.31692v1…