DF-TransUNet: A novel TransUNet model of pixel level classification for cardiac MR image segmentation

Abstract

A crucial step in the intelligent healthcare system is the automatic analysis of medical images by machines and processing them accordingly, particularly in disease diagnosis, as it provides accurate anatomical structure information for subsequent treatment. This process provides precise anatomical data vital for the subsequent treatments. There is a problem of uneven intensity distribution and fuzzy boundaries etc. in medical images, which creates a great problem in the segmentation task. To cope with this problem, an improved TransUNet structure is introduced in this paper. This method is based on the TransUNet framework, and adds a pixel level classification module on the basis of TransUNet segmentation. This module can further classify the boundary parts of the mask to be segmented at the pixel level to achieve more accurate segmentation results. The method in this paper effectively reduces the classification errors associated with pixels near the boundaries of various masks within the MR image. In particular, The pixel classification module aims to learn the category of pixels near the mask boundary, and then enhance the original segmentation results through pixel level adjustments. To validate the effectiveness of this model, a series of experiments were conducted using the 2017 MICCAI Automated Cardiac Diagnostic Challenge (ACDC) dataset. The results show that with an average Dice coefficient of 90.55% and a Hausdorff distance of up to 2.23 mm, the proposed approach achieves a commendable segmentation performance. Code and models are available at https://github.com/laodeyip/DF-TransUNet.