Improving Malaria diagnosis through interpretable customized CNNs architectures.

Area of Science:

• Medical diagnostics
• Artificial intelligence in healthcare
• Parasitology

Background:

• Malaria remains a critical global health issue, particularly in regions with high mosquito populations.
• Current diagnostic methods, such as manual microscopy of blood samples, are time-consuming, prone to errors, and require expert personnel.
• There is a pressing need for more efficient and accurate malaria detection tools.

Purpose of the Study:

• To develop and evaluate advanced deep learning models for improved malaria diagnosis.
• To compare the performance of custom convolutional neural networks (CNNs) against established transfer learning models.
• To identify the most effective AI model for rapid and accurate malaria parasite detection.

Main Methods:

• Development and implementation of three customized CNN architectures: Parallel convolutional neural network (PCNN), Soft Attention Parallel Convolutional Neural Networks (SPCNN), and Soft Attention after Functional Block Parallel Convolutional Neural Networks (SFPCNN).
• Evaluation of model performance using metrics including precision, recall, F1 score, accuracy, and Area Under the Receiver Operating Characteristic Curve (AUC).
• Comparison against various transfer learning algorithms (VGG16, ResNet152, MobileNetV3Small, EfficientNetB6, EfficientNetB7, DenseNet201, Vision Transformer (ViT), Data-efficient Image Transformer (DeiT), ImageIntern, and Swin Transformer v1/v2).
• Assessment of model interpretability using feature activation maps, Gradient-weighted Class Activation Mapping (Grad-CAM), and SHapley Additive exPlanations (SHAP).

Main Results:

• The Soft Attention Parallel Convolutional Neural Networks (SPCNN) model demonstrated superior performance, achieving high scores across all evaluation metrics (e.g., 99.38% precision, 99.37% recall, 99.95% AUC).
• SPCNN significantly outperformed all tested transfer learning models and other custom CNNs in diagnostic accuracy and speed.
• SPCNN exhibited the fastest testing time (0.00252 s) among the developed CNNs, indicating high computational efficiency.
• Interpretability analyses confirmed the robustness and effectiveness of the SPCNN model in identifying malaria parasites.

Conclusions:

• The developed SPCNN model represents a significant advancement in malaria parasite diagnosis, surpassing traditional manual microscopy.
• This AI-driven approach offers a highly accurate, rapid, and computationally efficient solution for malaria detection.
• The study underscores the potential of advanced deep learning technologies in developing effective tools for malaria prevention and control.