Shuang Zhang, Xianyu Zeng
BAR-UNet achieves Dice 0.8798 ±0.0024 and IoU 0.7913 ± 0.0029 on Kvasir-SEG, outperforming all compared methods. Zero-shot evaluation on CVC-ClinicDB, CVC-ColonDB, and ETIS-Larib yields Dice/IoU of 0.831/0.748, 0.776/0.682, and 0.728/0.631, respectively. Ablation studies, boundary metrics (F1^bd, MAE, HD95), paired Wilcoxon tests, and MACL sensitivity analyses confirm that the boundary head and MACL are complementary.
INTRODUCTION: Automatic detection of skin lesions in dermoscopic images remains challenging due to large intra-class variation, low-contrast boundaries, and severe foreground-background imbalance.
METHODS: We propose FD-YOLO-Skin, a frequency-domain enhanced YOLOv8 framework for single-class micronucleus lesion detection. FD-YOLO-Skin introduces (i) a Frequency-Domain Multi-Scale Feature Fusion (FMSFF) module in the neck to fuse low-frequency shape cues with high-frequency texture details via FFT/IFFT-based multi-branch spectral processing, and (ii) a Frequency-Domain Contrastive Learning (FDCL) module on the backbone that applies spectral augmentations and a contrastive objective to improve feature robustness under complex backgrounds.
RESULTS: On the ISIC-Style Micronucleus Lesion Detection Benchmark (ISIC-MLD; 10,015 de-identified dermoscopic images), FD-YOLO-Skin achieves an mAP@0.5 of 0.990 ± 0.003 (95% CI: [0.986, 0.994]) and an mAP@0.5:0.95 of 0.905 ± 0.006 on the held-out test split, with precision and recall above 0.97. Ablations show that FMSFF mainly improves recall for small or low-contrast lesions, whereas FDCL reduces false positives and improves precision relative to aggressive spatial-domain augmentation alone.
DISCUSSION: Explicit frequency-domain multi-scale fusion and contrastive regularization improve single-class skin lesion detection with modest computational overhead. Source code, preprocessed dataset splits, and model weights are available at https://anonymous.4open.science/r/skin2-B816/.