MedicalPatchNet: A Patch-Based Self-Explainable AI Architecture for Chest X-ray Classification

Interpretability doesn't require sacrificing accuracy: MedicalPatchNet achieves nearly identical diagnostic performance (AUROC 0.907) to a standard deep learning model (0.908) while being inherently interpretable, demonstrating that explainability and accuracy are not mutually exclusive in medical AI.

Doctors can understand why the AI made its decision: Unlike "black box" AI systems that require specialized explanation techniques, MedicalPatchNet shows which specific regions of a chest X-ray contributed to its diagnosis through visual highlighting, making it accessible to clinicians without AI expertise.

Better explanations catch AI mistakes: The model's localization accuracy for identifying disease locations was significantly higher (48.5% hit-rate) compared to post-hoc explanation methods like Grad-CAM (37.6%), suggesting that transparent AI is less likely to rely on misleading patterns that don't reflect actual medical findings.

Transparency builds clinical trust: By providing explicit, understandable explanations for each diagnostic decision, MedicalPatchNet addresses a major barrier to AI adoption in hospitals—clinicians' concerns about trusting "black box" systems with patient care.

Reproducibility supports responsible deployment: The authors released code and training scripts publicly, enabling other researchers and institutions to verify the system's safety and adapt it for their own clinical settings rather than treating it as proprietary technology.