Hybrid CNN-VIT Model for Automated Multi-Stage Malaria Parasite Classification in Microscopy Images

Abstract

The diagnosis of malaria continues to be a big problem for treatment and surveillance of the disease in a timely manner, especially in places with limited resources where there are no expert microscopists. In this research, a new method that uses deep learning is presented combining EfficientNetB0-based convolutional neural network (CNN) with a Vision Transformer (ViT) in a very effective way to the automatic multi-stage malaria parasite classification from blood smear microscopy images. The local morphological features are extracted hierarchically by the CNN component, whereas the global spatial relationships and the long-range contextual dependencies within the infected cells are modeled by the ViT module enabling the improvement in the discrimination of the visually similar parasite stages. To increase the model’s generalization between the different datasets and staining variabilities to even further extent, contrastive self-supervised learning (CSSL) has been integrated during the model training alongside morphology-aware data augmentation, so that the model can learn feature representations that are stage-consistent and stain-invariant. The results of the experiments reveal that the hybrid CNN–ViT model is a better performer than conventional deep learning architectures by a large margin, thus attaining 99.1% accuracy, 98.9% precision, 99.0% recall, and an F1-score of 98.9% on the validation set. These results underscore the integration of convolutional feature extraction and transformer-based global attention in the automation of malaria parasite analysis and point out that the proposed framework provides a trustworthy and scalable alternative to manual microscopy in clinical and field settings.