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

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Computer-assisted diagnosis

Background:

• Accurate medical image segmentation is vital for intelligent healthcare systems and disease diagnosis.
• Challenges in medical imaging include uneven intensity distribution and fuzzy boundaries, hindering segmentation tasks.
• Existing methods struggle with precise delineation of anatomical structures.

Purpose of the Study:

• To introduce an improved TransUNet structure for more accurate medical image segmentation.
• To address the limitations of uneven intensity and fuzzy boundaries in MR images.
• To enhance the segmentation of cardiac structures in medical scans.

Main Methods:

• An improved TransUNet framework incorporating a pixel-level classification module was developed.
• This module specifically targets and refines the classification of pixels near mask boundaries.
• The approach enhances original segmentation results through pixel-level adjustments.

Main Results:

• The proposed method significantly reduces classification errors near mask boundaries in MR images.
• Experiments on the 2017 MICCAI Automated Cardiac Diagnostic Challenge (ACDC) dataset were conducted.
• Achieved an average Dice coefficient of 90.55% and a Hausdorff distance of 2.23 mm.

Conclusions:

• The improved TransUNet model demonstrates commendable segmentation performance for cardiac MR images.
• The pixel classification module effectively enhances segmentation accuracy, particularly at boundaries.
• This approach contributes to more precise anatomical data for subsequent medical treatments.