SML: Improving Colorectal Polyp Detection Accuracy via Multi-Scale Annotations
Frontiers in Medicine, cilt.1, sa.1, ss.1-15, 2026 (Hakemli Dergi)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 1 Sayı: 1
- Basım Tarihi: 2026
- Dergi Adı: Frontiers in Medicine
- Sayfa Sayıları: ss.1-15
- Atatürk Üniversitesi Adresli: Evet
Özet
Object detection performance is jointly determined by model capacity and the quality of supervised data. Conventional single-scale bounding-box annotations provide only one spatial view per instance and may under-represent semantic variability, especially in data-scarce domains. We introduce Small–Medium–Large (SML) labelling, a simple dataset-level strategy that enriches supervision by annotating each object at three scales within the same image. Practically, images are duplicated and each duplicate receives scale-specific boxes (S, M, L), tripling the number of bounding boxes without adding new imagery. Unlike data-augmentation techniques that expand datasets by transforming raw pixels, SML keeps the imagery fixed and improves only the supervision (i.e., the quality and semantic depth of labels). On a custom COCO-format dataset, SML consistently improves accuracy across detectors. Relative to conventional M-only labelling, SML yields gains of 6–7% mAP@[.5,.95] and larger increases at stricter IoU (mAP@.75) for Faster R-CNN (R101/X101-FPN) and YOLO (v5/v7). Bene-fits are observed across all evaluated scales and detector architectures, indicating that multi-scale supervision provides complementary spatial information during object localization. For domain-specific verification, the pro-posed method was further applied to the Kvasir-SEG colorectal polyp detection dataset, where SML achieved 8% mAP improvement, demonstrating its robustness and applicability to medical imaging tasks. Although the medical domain provides a strong validation of its effectiveness, the proposed method remains broadly applicable to other visual domains such as agriculture, autonomous driving, and industrial inspection, where dataset expansion is costly but annotation refinement is feasible.