SDNAL-Seg: multi-scale selective downsampling and non-adjacent layers guidance for medical image segmentation

Medical image segmentation is dedicated to the precise extraction and delineation of specific anatomical structures, lesion areas, or regions of interest from medical images. Traditional approaches often adopt an encoder-decoder architecture, where direct downsampling operations in the encoder part inevitably introduce distortions that degrade segmentation performance. To address this challenge, we propose a novel modular design that integrates multi-scale feature extraction and downsampling functions into the encoder, forming a unified processing unit–the Multi-Scale Selective Downsampling (MSSD) Module. MSSD employs a synchronous downsampling and path selection mechanism to minimize distortion-induced effects on segmentation outcomes caused by downsampling. In addition, to achieve semantic information fusion between the encoder and decoder through skip connections, an enhanced module named the Non-Adjacent Layers Guidance (NALG) Module is designed. This module leverages guidance from both higher-level and lower-level layers to facilitate the integration of contextual information for boosting segmentation accuracy. For the decoder, we propose a Dual Dilated Fusion Attention (DDFA) module. Experimental results demonstrate that the proposed approach effectively balances performance and model complexity, making it well-suited for high-accuracy medical image segmentation tasks without requiring excessive parameters. The source code will be available at https://github.com/Linqb11/SDNAL-Seg.git.