MILCA: Malaria Parasite Detection from Sample-Level Weak Labels

Abstract

Malaria diagnosis requires the inspection of multiple image fields per sample. Training vision models for malaria parasite detection typically requires large numbers of expert-provided bounding boxes, which are costly to obtain and often impractical in real-world deployments. We introduce MILCA, a weakly supervised object detection framework that learns parasite localization from sample-level diagnostic labels, which are routinely recorded in clinical practice. MILCA combines Multiple Instance Learning (MIL) for sample classification with an iterative Class Activation (CA) Mapping procedure that yields coarse parasite pseudo-labels, which are further enriched with hard negatives from parasite-free samples. These pseudo-labels enable training a detector without any manual bounding-box supervision. Experiments on multiple microscopy datasets show that MILCA achieves reliable detection and counting performance under fully weak supervision, and that fine-tuning with only a small fraction of expert annotations provides substantial additional gains, outperforming supervised and pseudo-labeling baselines under the same or lower annotation budgets. By converting coarse, sample-level clinical labels into effective object-level supervision, MILCA provides a label-efficient route toward automated malaria parasite detection and a general approach for weakly supervised blood film analysis.