Hybrid Transformer Encoder-Decoder Framework with Collective Deep Learning for Semantic Skin-Lesion Segmentation

Abstract

Accurate skin-lesion segmentation remains difficult because dermoscopic images often contain low contrast, irregular borders, hair, illumination variation, and heterogeneous lesion appearance. This study proposes a hybrid Transformer encoder-decoder framework with collective deep learning to combine global contextual modeling with local boundary recovery and complementary prediction fusion. The pipeline uses Wiener filtering and dynamic histogram equalization, patch embedding, hierarchical Swin/Transformer encoding, skip-connected decoding, attention refinement, and normalized validation-based fusion of complementary prediction branches. The framework is evaluated on an internal 70%/15%/15% split of the 2,594 annotated ISIC 2018 Task 1 images using Dice, IoU, accuracy, sensitivity, specificity, computational profiling, component ablation, and descriptive multi-run variability. The proposed model achieved 95.28% Dice, 91.74% IoU, 97.88% accuracy, 95.11% sensitivity, and 98.14% specificity. The available ablation results show progressive gains from Wiener filtering, dynamic histogram equalization, hierarchical Transformer encoding, and attention refinement. The results support the value of integrating image enhancement, hierarchical contextual encoding, boundary-aware reconstruction, and collective prediction within a single lesion-segmentation pipeline. Because the present evidence is limited to an internal ISIC 2018 evaluation and does not include an independent external dataset, generalization claims are deliberately restricted to the evaluated benchmark.