<![CDATA[Skin cancer, particularly melanoma, has a high mortality rate, necessitating reliable early detection based on dermoscopic images. Accurate lesion segmentation is a crucial preprocessing step prior to deep learning–based classification. However, challenges remain in preserving lesion boundary details, handling variations in lesion size, and addressing limited data availability and hyperparameter optimization. This study proposes and evaluates various configurations of a hybrid VGG–SegNet model for skin lesion segmentation using the ISIC 2017 and ISIC 2018 datasets. The methodology includes image and ground truth validation, VGG16-based normalization, input image resolution variations (128×128, 160×160, and 256×256), and training–validation data splits of 50:50 and 70:30. The model is fine-tuned using skip connections and a multi-stage training scheme with a combination of Binary Cross-Entropy, Dice, and Focal Tversky Loss. The best-performing model achieves a Dice Coefficient of 0.899 and an Intersection over Union of 0.8177 on the validation set, demonstrating precise and efficient lesion segmentation.Keywords: Skin Cancer Segmentation; VGG-SegNet; ISIC; Fine Tuning AbstrakKanker kulit, khususnya melanoma, memiliki tingkat mortalitas tinggi sehingga memerlukan deteksi dini berbasis citra dermoskopi yang andal. Segmentasi lesi yang akurat merupakan tahap penting sebelum proses klasifikasi berbasis deep learning. Namun, tantangan masih muncul pada detail tepi lesi, variasi ukuran, serta keterbatasan data dan optimasi hyperparameter. Penelitian ini mengusulkan dan mengevaluasi berbagai konfigurasi model hibrida VGG–SegNet untuk segmentasi lesi kulit menggunakan dataset ISIC 2017 dan ISIC 2018. Tahapan meliputi validasi citra dan ground truth, normalisasi berbasis VGG16, variasi resolusi citra (128×128, 160×160, 256×256), serta pembagian data latih dan validasi dengan rasio 50:50 dan 70:30. Model di fine tuning menggunakan skip connection dan skema pelatihan bertahap dengan kombinasi Binary Cross-Entropy, Dice, dan Focal Tversky Loss. Model terbaik mencapai Dice Coefficient 0,899 dan IoU 0,8177 pada data validasi, menunjukkan segmentasi yang presisi dan efisien.Kata kunci: Segmentasi Kanker Kulit; VGG–SegNet; ISIC; Fine Tuning]]>
