Facial Pigmented Lesion Classification with SwiGLU-Enhanced Swin Transformer V2
Mediterranean Conference on Medical and Biological Engineering and Computing (2026), Siena, İtalya, 14 - 17 Eylül 2026, cilt.147, ss.341-356, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Cilt numarası: 147
- Doi Numarası: 10.1007/978-3-032-39674-7_23
- Basıldığı Şehir: Siena
- Basıldığı Ülke: İtalya
- Sayfa Sayıları: ss.341-356
- İstanbul Üniversitesi Adresli: Evet
Özet
Automated classification of facial pigmented lesions from dermoscopic images is a clinically important yet challenging problem due to the high visual similarity between malignant and benign lesions, severe class imbalance, and scarcity of validated training data. This paper presents our approach to the MEDICON 2026 Scientific Challenge, which targets seven-class differential diagnosis of facial pigmented lesions. Our key contributions are twofold. First, we curated an extended, dermatologist-validated dataset by merging the official MEDICON 2026 training images with publicly available external sources (ISIC Archive and HAM10000), followed by GPT-4 Vision-assisted subclass disambiguation and manual clinical validation by a board-certified dermatologist who reviewed and removed any suspicious or incorrectly labelled samples—yielding approximately 400–500 validated images per class. Second, we propose a classification pipeline based on Swin Transformer V2 (pre-trained on ImageNet-22K) with a custom two-layer SwiGLU classification head, trained with Focal Loss, MixUp/CutMix augmentation, Exponential Moving Average (EMA) of weights, and 8-view Test-Time Augmentation (TTA) at inference. On our internal validation set, the model achieved a competition score of 87.95, with Macro-AUC of 0.972, Averaged Recall of 0.811, and Macro-F1 of 0.810. The official hidden test score was 61.65, revealing a distribution shift between the external-data-enriched training set and the clinical test set—a finding with important implications for future challenge design.