Improving Malaria diagnosis through interpretable customized CNNs architectures.

Malaria, which is spread via female Anopheles mosquitoes and is brought on by the Plasmodium parasite, persists as a serious illness, especially in areas with a high mosquito density. Traditional detection techniques, like examining blood samples with a microscope, tend to be labor-intensive, unreliable and necessitate specialized individuals. To address these challenges, we employed several customized convolutional neural networks (CNNs), including Parallel convolutional neural network (PCNN), Soft Attention Parallel Convolutional Neural Networks (SPCNN), and Soft Attention after Functional Block Parallel Convolutional Neural Networks (SFPCNN), to improve the effectiveness of malaria diagnosis. Among these, the SPCNN emerged as the most successful model, outperforming all other models in evaluation metrics. The SPCNN achieved a precision of 99.38 ± 0.21%, recall of 99.37 ± 0.21%, F1 score of 99.37 ± 0.21%, accuracy of 99.37 ± 0.30%, and an area under the receiver operating characteristic curve (AUC) of 99.95 ± 0.01%, demonstrating its robustness in detecting malaria parasites. Furthermore, we employed various transfer learning (TL) algorithms, including VGG16, ResNet152, MobileNetV3Small, EfficientNetB6, EfficientNetB7, DenseNet201, Vision Transformer (ViT), Data-efficient Image Transformer (DeiT), ImageIntern, and Swin Transformer (versions v1 and v2). The proposed SPCNN model surpassed all these TL methods in every evaluation measure. The SPCNN model, with 2.207 million parameters and a size of 26 MB, is more complex than PCNN but simpler than SFPCNN. Despite this, SPCNN exhibited the fastest testing times (0.00252 s), making it more computationally efficient than both PCNN and SFPCNN. We assessed model interpretability using feature activation maps, Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) visualizations for all three architectures, illustrating why SPCNN outperformed the others. The findings from our experiments show a significant improvement in malaria parasite diagnosis. The proposed approach outperforms traditional manual microscopy in terms of both accuracy and speed. This study highlights the importance of utilizing cutting-edge technologies to develop robust and effective diagnostic tools for malaria prevention.