A Medical image segmentation model with auto-dynamic convolution and location attention mechanism.

Background And Objective: Medical image segmentation is a technique used to identify and locate anatomical structures or diseased areas from medical images with high accuracy. Accurate image segmentation is crucial in medical applications such as clinical diagnosis, surgical planning, and treatment monitoring. It provides reliable quantitative information, which helps in making decisions. The current models used for medical image segmentation are good at capturing far-reaching and global context information. However, these models have limited resolution due to their high computing requirements. Additionally, many of the models lack important information between slices, which reduces overall performance.

Methods: To address these challenges, we introduced a new architecture: auto-dynamic convolution with location attention former(AD-LA Former). We have developed a novel method that utilizes auto-dynamic convolution and location attention mechanism to dynamically adapt to different data patterns. Our model incorporates an internal scaling layer to enhance the dynamics of training, integrates auto-dynamic convolution to comprehensively learn the different attention of convolution kernel, and introduces location attention to obtain more precise spatial location information and dependency.

Results: We evaluated our model against leading methods on popular medical segmentation datasets such as synapse, ISIC2017 and ISIC2018 datasets, and it has demonstrated better performance compared to other methods. On the synapse dataset, DSC can reach 83.48; on the ISIC2017 dataset, ACC, SE and SP can respectively reach 0.9703, 0.9865 and 0.9295. On the ISIC2018 dataset, ACC, SE and SP can respectively reach 0.9646, 0.9904 and 0.9124.

Conclusions: AD-LA Former can solve the problem of redundant information between channels and realize the ability to capture cross-channel information, so that more accurate and efficient segmentation results can be obtained.