Abstract

Malaria diagnosis in many endemic settings continues to depend on microscopic reading of thick and thin blood smears. Deep learning detectors now perform well on such images, but they are almost always trained and tested within a single smear type, and it remains unclear whether the features they learn on one preparation carry over to the other. The two preparations are not equivalent. Thin smears keep the red blood cells intact and preserve a cellular frame around the parasite, whereas thick smears lyse those cells and leave the parasite against a background of haemoglobin debris and staining artefacts. We therefore treat thick and thin smears as separate visual domains and train YOLO-based detectors under three regimes: thick-only, thin-only, and combined. Within its own validation domain the thick model reached an mAP50 of 0.866 and an mAP50-95 of 0.443, while the thin model reached 0.678 and 0.575. Per-class results were strong for leukocytes and trophozoites and noticeably weaker for gametocytes and difficult cases, and an early combined model raised recall but lowered precision and mAP. Our observations indicate that these numbers are best read as smear-specific baselines rather than as a direct measurement of transfer, and they offer preliminary evidence that preparation type behaves as a domain boundary. We are explicit that the confirmatory test, evaluating each model on the other preparation, has not yet been run, and we set out the protocol we intend to follow. On the present evidence we suggest that malaria detection models report performance separately for each smear type, and that transfer across preparations not be assumed without direct measurement.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_050.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=QPKDQiEvCn

BibTex

@InProceedings{DakPet_Does_MICCAISAT2026,
        author = { Daka, Peter Chundama AND Muzumala, Gabriel Malaizyo AND M’ule, Barbara Chanda AND Kabindama, Mbunji AND Ismaila, Lukman E.},
        title = { { Does Smear Preparation Type Constitute a Domain Boundary? Within-Domain Baselines and a Transfer Protocol for Malaria Microscopy } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17264},
        month = {pending},
        page = {pending}
}


back to top