Depth disentanglement strategy of latent space for medical image segmentation

In medical images, there are always numerous tiny structures such as narrow blood vessels and small cell nuclei. In most cases, the pathological information contained within these small structures can have a significant impact on the diagnosis of diseases. Many previous methods were unable to simultaneously consider the microscopic details and macroscopic global aspects of the images, making errors in image segmentation more likely. Therefore, the purpose of this paper is to propose a medical image segmentation strategy based on deep feature disentanglement. This strategy effectively disentangles the detailed features of local regions by leveraging small convolutions with strong inductive biases in the shallow layers. Simultaneously, it achieves effective modeling of global features by disentanglement channel features in the deep layers, taking into account regional consistency. In medical images, the presence of noise can lead to incorrect interpretations of pathological information. To address this, we perform deep feature mining at multiple stages, focusing on the finest details that are difficult to distinguish from noise in the image. We use spatial-channel attention mechanisms to model feature information at multiple stages and learn the differences between these two types of information. We conducted experiments on three different datasets, DRIVE, CHASE-DB, and MoNuSeg, which contain a large number of small segmentation structures, without any pre-training. The experimental results demonstrate that our network outperforms state-of-the-art medical image segmentation models significantly.