A Dual-Objective Contrastive Learning-Based Deep Learning Model for Malaria Diagnosis

Malaria is one of the most life-threatening infectious diseases worldwide, necessitating early and accurate diagnosis to ensure effective treatment. Traditional microscopic examination of blood smears is labor-intensive and prone to diagnostic inconsistencies. Additionally, many existing deep learning (DL) approaches employ conventional transfer learning (TL) strategies that fine-tune pretrained convolutional neural networks (CNN) or use them as fixed feature extractors, thereby constraining domain adaptability and increasing reliance on extensive labeled datasets. Therefore, this study introduces a self-supervised DL framework that integrates contrastive representation learning with TL for automated malaria diagnosis. The proposed model employs a dual-objective contrastive fine-tuning strategy that jointly optimizes supervised cross-entropy and self-supervised contrastive objectives. This hybrid training paradigm enforces semantic alignment between positive sample pairs while maximizing inter-class separation, thereby yielding highly discriminative embeddings. Multiple CNN backbones, including EfficientNetB7, Revitalized DenseNet Reloaded, ConvNeXt, and RegNetX, serve as the foundational architectures for implementing this strategy. The contrastive EfficientNetB7 attained the highest accuracy of 96.87%, outperforming all baseline models. Unlike conventional pipelines that rely on handcrafted preprocessing, segmentation, and multi-stage classification, the proposed framework delivers an end-to-end generalizable solution. This work provides a new perspective on the integration of self-supervised learning within TL to enhance diagnostic reliability and representation quality in medical image analysis.