Beyond Accuracy: Evaluating Efficiency, Robustness and Explainability in Deep Learning for Malaria Diagnosis

Lightweight models perform as well as larger ones: The study found no statistically significant performance differences between efficient, lightweight deep learning models and heavier counterparts on malaria diagnosis, suggesting that resource-constrained settings in sub-Saharan Africa don't need to sacrifice accuracy for computational efficiency.

Explainability methods are unreliable under real-world conditions: While CAM-based explanation methods successfully identified diagnostically relevant regions in clean images, none of the explainability methods remained robust when images were corrupted—a critical finding since clinical settings often involve imperfect image quality that could occur in practice.

Model confidence is a better warning signal than accuracy: The research revealed that model confidence degrades faster than actual accuracy when images are corrupted, meaning clinicians could use a drop in the model's confidence as a practical trigger to request human expert review, even when the model's predictions remain technically accurate.

Post-hoc explanations shouldn't be trusted as the sole basis for clinical decisions: The finding that explanation reliability degrades at noise levels "plausible in clinical practice" suggests that AI explanations alone are insufficient for clinical adoption—human expert validation remains essential, particularly in resource-limited settings where image quality may be variable.

Deployment strategy matters for responsible AI in healthcare: The paper supports using lightweight architectures in resource-constrained regions while emphasizing that explainability methods require further development before they can reliably support clinical decision-making in real-world conditions.