Addressing Class Imbalance in Skin Lesion Classification Using Swin Transformer: A Progressive Ablation Study on ISIC 2019
16th International Conference on Advanced Computer Information Technologies, ACIT 2026, Zlin, Çek Cumhuriyeti, 2 - 04 Eylül 2026, ss.1385-1389, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/acit69831.2026.11700101
- Basıldığı Şehir: Zlin
- Basıldığı Ülke: Çek Cumhuriyeti
- Sayfa Sayıları: ss.1385-1389
- Anahtar Kelimeler: class imbalance, dermoscopy, focal loss, logit adjustment, multi class classification, skin lesion classification, Swin Transformer
- Kocaeli Üniversitesi Adresli: Evet
Özet
Skin cancer is one of the most common types of cancer worldwide. Although there have been improvements in deep learning for medical imaging applications, it is still difficult to automate the process of diagnosing skin diseases due to the availability of insufficient data and class imbalance issues. We investigate this problem on the eight Class ISIC 2019 benchmark using a SwinV2 Tiny backbone and a four stage progressive ablation. Each stage adds one component, respectively, stochastic augmentation with MixUp, Exponential Moving Average (EMA) weight consolidation, Test Time Augmentation (TTA), and finally composite loss that merges focal gradient modulation with prior probability logit correction. The complete pipeline raises Macro F1 from 0.7411 to 0.8195 and the Matthews Correlation Coefficient (MCC) to 0.8194, achieving a macro AUC of 0.9662 and a macro sensitivity of 0.8276. Beyond these numerical gains, the pipeline provides significant improvement numerically, but more importantly, it allows for an increase in structural generalization abilities when dealing with rare lesions without the use of synthesized data.