clinical-ai

This study provides a rigorous and model-aware examination of the faithfulness and spatial reliability of Grad-CAM explanations for lung cancer CT image classification across both convolutional neural networks and Vision Transformer architectures. By systematically analyzing localization accuracy, perturbation-based faithfulness, and explanation consistency, the work reveals pronounced architecture-dependent disparities in how visual explanations align with true diagnostic evidence. While Grad-CAM often produces visually convincing heatmaps for convolutional models, these explanations can be spatially coarse or influenced by spurious correlations, raising concerns about shortcut learning and misleading interpretability. More critically, the analysis demonstrates that transformer-based models, despite strong predictive performance, exhibit a marked degradation in Grad-CAM reliability due to non-local attention mechanisms. Together, these findings underscore a central message, visually appealing explanations do not necessarily imply faithful model reasoning. The work highlights fundamental limitations of saliency-based XAI methods in high-stakes medical imaging and calls for more principled, model-aware interpretability approaches that can support genuinely trustworthy and clinically meaningful AI systems.

This work introduces LymphAware, a domain-aware bias disruption framework designed to improve the robustness and clinical reliability of lymphoma histopathology AI systems. Modern classifiers often rely on shortcut signals such as scanner-specific color distributions, staining variability, and slide-preparation artifacts, which artificially inflate in-domain accuracy but collapse under cross-acquisition domain shift. LymphAware explicitly disentangles disease-relevant morphological features from acquisition-sensitive nuisance factors through three complementary mechanisms, morphology-centric feature isolation, adversarial and orthogonality-based shortcut suppression, and cross-domain stability regularization. To further expose hidden shortcut dependencies, the framework incorporates artifact-shift perturbations that simulate realistic staining and scanner variability while enforcing counterfactual consistency during training. Extensive evaluation on a heterogeneous multi-source lymphoma benchmark demonstrates improved cross-domain generalization, stable behavior under hyperparameter variation, and attribution maps better aligned with pathology-relevant regions. While these explanations reflect associative alignment rather than formal causal inference, the findings underscore the necessity of representation-level shortcut disruption for building clinically trustworthy lymphoma diagnostic AI systems.