Abstract
Malaria’s burden falls hardest where reliable microscopy is scarcest, so a detector tuned in one context must stay safe when redeployed in another. This tempts a shortcut: carrying a recall-boosting loss, Quality-Guided Focal Loss (QGFL), unchanged from one detector to the next. We test it by transferring QGFL across three detectors, YOLOv8s, YOLOv11s and RT-DETR, on three malaria datasets, across 135 training runs. The shift towards recall is universal, but its effect on F1 is not. Aggressive re-weighting helps a minority and quietly harms the rest. What decides the outcome is one measurable property of the untouched baseline, its precision–recall balance, not its architecture or accuracy. Matched baselines with opposite balance moved in opposite directions, and the verdict survives a recall-weighted re-scoring. This yields a pre-deployment check, applying the aggressive loss only when the baseline is precision-dominant and flagging harmful transfer before retraining. The lesson reaches past this loss: reliability comes from matching the intervention to the baseline’s error profile, not from the strongest one. Code: https://github.com/ThabangIsaac1/qgfl-transfer-malaria.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AFRICAI_020.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=MiXjO7o8HJ
BibTex
@InProceedings{IsaTha_Harmful_MICCAISAT2026,
author = { Isaka, Thabang Fenge AND Wynne, Claire AND Courtney, Jane},
title = { { Harmful or Helpful? When Precision-Recall Balance Governs Safe QGFL Transfer in Malaria Detection } },
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}
}