Deep Learning Approaches for Malaria Diagnosis: A Comparative Study of Custom CNN and Transfer Learning Models in Blood Smear Analysis

Abstract

The Plasmodium genus of single-celled parasites is the cause of malaria. These parasites are spread to humans through the bite of an infected Anopheles mosquito. Particularly in sub-Saharan Africa, where it severely strains health systems and economics, the illness remains an essential public health issue. Effective diagnosis and treatment are crucial for controlling and eventually eliminating malaria. The microscopic analysis of blood smears is the conventional yet labour-intensive method for diagnosing malaria, demanding extensive expertise. Automated detection through deep learning presents a vital alternative, particularly crucial for sub-Saharan Africa. This study aims to meticulously compare the performance of a custom-designed Convolutional Neural Network (CNN) with five advanced transfer learning models, ResNet50, VGG19, InceptionV3, EfficientNet-B3, EfficientNet- B7, and YOLOv11m, in categorizing segmented red blood cell images for malaria detection. Our approach involves comprehensive image preprocessing, data augmentation, and the implementation of various models. The models were evaluated using a National Library of Medicine dataset based on various metrics, including F1 Score, Accuracy, Precision, Recall, Matthews Correlation Coefficient (MCC), AUC ROC, and AUC PR. The EfficientNet-B3 model emerged as the top performer, surpassing even the custom CNN with an impressive F1 Score of 98.12%, Accuracy of 98.08%, and an MCC of 96.15%, demonstrating its superior predictive power and reliability. YOLOv11m also showed strong performance with an F1 Score of 96.89%, Accuracy of 96.93%, and MCC of 93.91%, highlighting its efficiency for real-time applications. Although the custom CNN did not outperform the advanced models, it still exhibited commendable performance, underscoring the potential of tailored architectures. The results of this study demonstrate the great potential that deep learning methods have to improve the precision of malaria diagnosis, providing notable benefits to the healthcare systems, especially for regions most severely impacted by the disease.