Thoracic and Abdominal Medical Image Segmentation Based on Dual-Branch Feature Interaction and Cross-Sample Feature Fusion

Key Points

• This research aims to improve the accuracy of thoracic and abdominal medical image segmentation by addressing common errors caused by artifacts and blurry boundaries.
• Developed a semantic segmentation network called MQUNet using a dual-branch feature interaction mechanism.
• Utilized a CNN-Transformer architecture to extract local and global features from medical images.
• Employed cross-sample feature fusion to dynamically incorporate stable anatomical prior features during the segmentation process.
• Achieved a Dice coefficient of 92.53% on the cardiac tumor dataset (ACDC).
• Achieved a Dice coefficient of 85.17% on the multi-organ abdominal dataset (Synapse).
• Demonstrated significant advantages in image segmentation over existing methods.