Towards Transparent Diagnostics: Investigating Architectural Trade-offs and Explainability in Malaria Detection

Efficiency vs. Accuracy Trade-offs Matter for Deployment: MobileNetV2 achieved 96.85% accuracy while being 8 times smaller (8.49 MB) and 10 times faster (1.35 ms inference) than the best-performing ResNet18+TTA model (97.96% accuracy, 13.32 ms), highlighting that the "best" model depends on real-world constraints like available computing resources in resource-limited settings where malaria is prevalent.

Explainability Tools Reveal Model Reliability: The paper used three different explainability methods (GRAD-CAM, SHAP, and LIME) to evaluate whether models were making predictions for the right reasons, suggesting that agreement or disagreement between these explanation methods can help clinicians assess whether they should trust a model's diagnosis in practice.

Model Optimization Doesn't Always Sacrifice Performance: Pruning the ResNet18+TTA model (removing unnecessary components) actually improved accuracy slightly while reducing inference time, demonstrating that responsible AI deployment can achieve both efficiency and reliability rather than forcing a choice between them.

Multiple Metrics Beyond Accuracy Are Essential: The paper reported not just accuracy (97.96%) but also AUC (0.996) and model size/speed, recognizing that a complete picture of model performance requires examining different dimensions—important for medical applications where false negatives could have life-or-death consequences.

Transparency Supports Clinical Adoption: By providing visual explanations of how models make malaria detection decisions, explainability methods help bridge the gap between "black box" AI and clinical practice, potentially increasing doctor confidence in using these tools to augment expert diagnosis in resource-constrained settings.