Towards Field-Ready AI-based Malaria Diagnosis: A Continual Learning Approach

TL;DR

This paper explores using continual learning techniques to improve the robustness and generalization of AI-based malaria diagnosis systems across different clinical sites, aiming for practical field deployment.

Contribution

It introduces a domain-incremental learning framework for malaria CAD models and evaluates multiple CL strategies on real multi-site data, demonstrating improved performance.

Findings

Rehearsal-based CL methods significantly enhance model robustness.

Continual learning reduces performance degradation across sites.

The approach supports development of deployable malaria diagnostic tools.

Abstract

Malaria remains a major global health challenge, particularly in low-resource settings where access to expert microscopy may be limited. Deep learning-based computer-aided diagnosis (CAD) systems have been developed and demonstrate promising performance on thin blood smear images. However, their clinical deployment may be hindered by limited generalization across sites with varying conditions. Yet very few practical solutions have been proposed. In this work, we investigate continual learning (CL) as a strategy to enhance the robustness of malaria CAD models to domain shifts. We frame the problem as a domain-incremental learning scenario, where a YOLO-based object detector must adapt to new acquisition sites while retaining performance on previously seen domains. We evaluate four CL strategies, two rehearsal-based and two regularization-based methods, on real-life conditions thanks to a multi-site clinical dataset of thin blood smear images. Our results suggest that CL, and rehearsal-based methods in particular, can significantly improve performance. These findings highlight the potential of continual learning to support the development of deployable, field-ready CAD tools for malaria.