Enhanced Cross-stage-attention U-Net for esophageal target volume segmentation

Area of Science:

• Medical Imaging
• Radiotherapy Planning
• Computational Anatomy

Background:

• Accurate segmentation of target volumes and organs at risk (OAR) is crucial for effective radiotherapy.
• Esophageal segmentation in CT images is challenging due to complex anatomy and low tissue contrast.
• Existing methods struggle with precise localization and scaling of the esophagus.

Purpose of the Study:

• To propose an Enhanced Cross-stage-attention U-Net for accurate segmentation of esophageal GTV and CTV.
• To address the difficulties in segmenting the esophagus in CT images for radiotherapy.
• To improve the efficiency and accuracy of radiotherapy planning through better esophageal segmentation.

Main Methods:

• Developed a novel Enhanced Cross-stage-attention U-Net architecture.
• Incorporated a principal component analysis module for initial feature extraction.
• Utilized a cross-stage feature fusion model with Wide Range Attention (WRA) and Small-kernel Local Attention (SLA) units.
• Implemented an Inverted Bottleneck unit with a global frequency response layer for feature map enhancement.

Main Results:

• The proposed method achieved competitive segmentation results for esophageal GTV and CTV.
• Mean Surface Distance (MSD) for GTV: 2.83 mm; for CTV: 5.26 mm.
• Hausdorff Distance (HD) for GTV: 11.79 mm; for CTV: 16.22 mm.
• Dice Coefficient (DC) for GTV: 72.45%; for CTV: 71.06%.

Conclusions:

• The Enhanced Cross-stage-attention U-Net demonstrates superior performance in esophageal GTV and CTV segmentation.
• Reconstruction of skip concatenation in U-Net architecture improved segmentation accuracy.
• The proposed network offers a more effective solution for esophageal segmentation in radiotherapy planning.