A COMPARATIVE EVALUATION OF DEEP LEARNING ARCHITECTURES FOR AUTOMATED MALARIA PARASITE DETECTION IN BLOOD SMEAR IMAGES

Abstract

Malaria continues to pose as the most prevalent global health problem with a particular predominance in developing countries lacking modern diagnosing mechanisms. Correct identification and efficient diagnosis of the malaria parasite are extremely crucial for treatment and containment of the disease. The research analyzes performance of advanced deep learning architectures to perform the classification of infected and uninfected blood cells using microscopic images. Five state-of-the-art models-EfficientNet-B2, Vision Transformer Small (ViT-S/16), ResNet-152, DenseNet-201, and Swin Transformer Base-were selected and tested after training on 27,558 labeled microscopic blood images obtained from National Institute of Health. The experiments used transfer learning and partial fine-tuning with BCE With Logits Loss and AdamW with cosine annealing for optimal classification of images. Swin Transformer Base delivered the highest test accuracy of 97.73% and an AUC value of 0.9969. The best performance on accuracy vs efficiency trade-off was achieved by EfficientNet-B2 with an accuracy and AUC of 97.10% and 0.9962, respectively. Transformer models produced sensitivities and specificities greater than 96%. The comparative analysis helps identify current strengths of deep learning for medical image classification and assists practitioners in selecting models in low-resource medical settings.