Trustworthy deep learning for malaria diagnosis using explainable artificial intelligence.

Area of Science:

• Medical Diagnostics
• Artificial Intelligence
• Computational Biology

Background:

• Malaria diagnosis faces challenges with traditional methods like microscopy, rapid diagnostic tests (RDTs), and polymerase chain reaction (PCR) due to limitations in scalability, sensitivity, and expertise.
• Automated diagnostic strategies are crucial for improving malaria detection, especially in resource-limited healthcare settings.

Purpose of the Study:

• To investigate the efficacy of deep learning models for automated malaria detection using blood smear images.
• To evaluate and compare the performance of various convolutional neural network (CNN) architectures.
• To enhance model interpretability and clinical trust through explainable artificial intelligence (XAI) techniques.

Main Methods:

• Empirical evaluation of four CNNs (MobileNetV2, VGG19, InceptionV3, ResNet18) on malaria blood smear images.
• Fine-tuning of advanced hybrid architectures (Xception, Inception-ResNetV2) on a large dataset (27,090 images).
• Validation of model robustness on an independent dataset with varied staining and imaging conditions.
• Application of XAI techniques (Grad-CAM, LIME, SHAP) for model interpretability.

Main Results:

• ResNet18 achieved the highest F1-score (96.33%) among the initial CNNs.
• Xception and Inception-ResNetV2 attained approximately 98% classification accuracy on validation and test sets.
• High accuracy (97-98%) was maintained on an independent dataset, demonstrating generalization.
• XAI techniques provided spatial, superpixel, and pixel-level transparency into model decisions.

Conclusions:

• Deep learning models, particularly hybrid architectures like Inception-ResNetV2, offer a highly accurate and scalable solution for malaria detection.
• Explainable AI techniques significantly enhance the interpretability and clinical trustworthiness of these automated diagnostic systems.
• The proposed AI framework addresses limitations of traditional methods, paving the way for improved malaria diagnostics in resource-limited settings.