Transformer-assisted broad learning for hybrid intelligence-based skin cancer segmentation

With the rise of Transformer architectures, deep learning applications have gradually shifted from traditional convolutional neural networks to Transformers based on self-attention mechanisms. In tasks such as image classification, segmentation, and detection, pure Transformer models (such as ViT, Swin Transformer, PVT) and hybrid architectures (such as ConViT, BoTNet) have become mainstream. At the same time, Broad Learning Systems (BLS) exhibit unique advantages in real-time scenarios with limited resources and small sample sizes, thanks to efficient training, strong dynamic scalability, and global optimization mechanisms. However, BLS has shortcomings in feature design, computational complexity, representation capability, and incremental learning stability. Transformer architectures effectively address these deficiencies through self-attention mechanisms, hierarchical structures, and parameter sharing. This paper proposes a hybrid network structure, VIWDNet, that combines the advantages of Transformers and BLS. The network retains the efficient training characteristics of BLS and incorporates the powerful feature abstraction capabilities of Transformers, making it suitable for skin lesion segmentation tasks in medical image analysis-a key step in early skin cancer screening and diagnosis. Experiments on four public medical image datasets (ISIC 2016, ISIC 2017, ISIC 2018, and PH2) show that VIWDNet improves Dice and mIoU metrics by 1.48-2.23% over the most advanced existing models, providing reliable technical support for computer-aided dermatological diagnosis.