A KMITL team combined YOLOv4-tiny, U-Net and DINOv2 to detect, segment and classify Candida in microscopy. Reported performance was high across detection, segmentation, binary and four-class tasks, with five-fold cross-validation.
Key findings
- Detection achieved mAP@50 0.908 and PR-AUC 0.949; segmentation Dice was 0.930 and IoU 0.874. Binary accuracy reached 0.980–0.988 and four-class F1 reached 0.977.
Why this matters globally
Fast, consistent fungal identification could support laboratories with limited expertise, particularly for immunocompromised patients.
Thai researcher contribution
Ten KMITL authors built the pipeline from image preparation through classification, led by corresponding author Veerayuth Kittichai.
Limitations to consider
Dataset size and diversity are unclear in the abstract. Internal cross-validation may be optimistic, and prospective workflow, time, cost and patient-impact testing are absent.