ADF-Net: A novel adaptive dual-stream encoding and focal attention decoding network for skin lesion segmentation

Automatic segmentation of lesion areas in dermoscopic images is a crucial step in computer-aided skin lesion examination and diagnosis systems. Efficient and accurate skin lesion segmentation benefits the quantitative analysis of diseases, such as melanoma, dermatofibroma, and seborrheic keratosis and so on. However, in practical clinical diagnosis, some lesion areas exhibit large-scale changes, fuzzy and irregular boundaries, and low contrast between the lesion and the background, leading to potential segmentation errors. To overcome this difficulty, we propose a novel network called ADF-Net, composed of a multi-stage dual-stream hybrid framework (MDHF) based on Transformer and a convolutional neural network is well designed to achieve comprehensive integration of coarse-grained and fine-grained feature representations. This framework, in combination with the adaptive feature fusion (AFF) module, enables efficient and adaptive integration of both global and local feature information. Additionally, a focal attention decoder (FAD) is proposed to suppress background noise, focusing on the target area, and promoting the fusion of encoder features and high-level features. Finally, we conduct extensive experiments on four public datasets, including ISIC 2018, ISIC 2017, ISIC 2016, and PH2. The test sets of them consist of 379, 600, 518, and 100 images, respectively. The results demonstrate that our ADF-Net outperforms other state-of-the-art methods in all four commonly used evaluation metrics, with Jaccard Index values of 84.52%, 78.92%, 87.44%, and 93.26% respectively. Furthermore, the computational complexity of the model is only 8.29G GFLOPs, and the inference time is only 20.6 ms, which is of significant importance for exploring clinical applications.