A Systematic Review of Transfer Learning Approaches for Malaria Diagnosis Using Red Blood Cell Imaging

Malaria remains one of the leading global health burdens, particularly in low-resource regions where access to reliable diagnosis is limited. Conventional microscopy is labor-intensive and dependent on skilled technicians, making it an ideal target for automation. To address this, recent studies have applied deep learning (DL) and transfer learning (TL) techniques to automate malaria diagnosis using red blood cell (RBC) images, achieving remarkable progress in both accuracy and deployment potential. This review consolidates and analyzes 25 peer-reviewed studies that explore various AI-driven malaria detection approaches, focusing on their contributions in model development, data utilization, and performance optimization. It provides a comprehensive synthesis of methods, challenges, and future directions in the field. The analysis employed a structured comparative method by extracting and summarizing key aspects from each study, including datasets, preprocessing techniques, transfer learning strategies, classification models, evaluation metrics, limitations, and recommendations. Tables were constructed to facilitate cross-study comparisons. The results show that most studies achieved high classification accuracy (often above 95%), particularly those using pretrained CNN architectures like VGG16, ResNet, and DenseNet. Several studies extended to species-level or stage-specific classification using multi-class models or transformer-based frameworks. Preprocessing strategies such as color normalization, segmentation, and augmentation were essential for boosting model performance. However, issues like class imbalance, dataset bias, annotation inconsistency, and lack of real-world validation persist across studies. Challenges in generalizability and computational scalability remain key barriers to clinical deployment. Future directions include using GANs for data balancing, adopting domain adaptation and federated learning, and embedding models into mobile or cloud-based diagnostic platforms. In conclusion, while deep learning approaches for malaria detection are technically mature and highly accurate under experimental conditions, broader clinical integration requires robust validation, dataset diversification, and interdisciplinary collaboration.