Att-Next for skin lesion segmentation with topological awareness

Skin lesion segmentation is crucial for the early detection and accurate diagnosis of dermatological conditions, as precise boundary delineation enables better identification of lesion features. While Convolutional Neural Networks (CNNs) and hybrid CNN-Attention models have achieved notable success in this task, they often struggle to segment fine-grained lesion boundaries and suppress irrelevant tumor-like artifacts. They also tend to neglect topological features, which are crucial for accurately identifying complex lesions. To address these limitations, we propose a novel hybrid model that integrates ConvNeXt blocks with self-attention mechanisms. The model is also enhanced by a topological loss combined with Binary Cross Entropy (BCE) loss. This approach enables the model to better capture both local and global context, accurately delineate lesion boundaries, and suppress irrelevant regions, all without relying on a pre-trained backbone. Our method is evaluated on four publicly available skin lesion datasets: ISIC 2016, ISIC 2018, HAM10000, and PH2. Performance is assessed using segmentation metrics such as the Dice coefficient and Jaccard index. Experimental results demonstrate that the proposed model outperforms state-of-the-art (SOTA) methods, including MISSFormer, Swin-UNet, LeViT-UNet, FAT-Net, Att-UNet, DoubleU-Net, DeepLabV3 and TransUNet. Notably, the model achieves a Jaccard index of 0.8529 and a Dice coefficient of 0.913 on the ISIC 2018 dataset, surpassing the performance of given SOTA models in boundary delineation and tumor-like region suppression. These results highlight the potential of our hybrid ConvNeXt-Attention model with topological loss to improve lesion segmentation accuracy, which would lead to more effective and precise dermatological diagnoses.