BA-UNet: A Boundary Augmented Segmentation Network for Cervical Cancer Radiotherapy.

Area of Science:

• Medical Imaging
• Radiotherapy
• Computational Biology

Background:

• Accurate segmentation of Clinical Target Volume (CTV) and Organs At Risk (OARs) is crucial for effective cervical cancer radiotherapy.
• Challenges include low tissue contrast of CTV and complex gastrointestinal organ geometry, hindering precise delineation.

Purpose of the Study:

• To develop an advanced segmentation framework, BA-UNet, that enhances geometric consistency for improved radiotherapy planning.
• To address limitations of existing methods in delineating indistinct boundaries and complex structures.

Main Methods:

• Proposed BA-UNet framework with a Boundary-Infused Feature Aggregation (BIFA) module to preserve high-frequency boundary information.
• Introduced a Boundary-Aware Curvature (BAC) loss function using Hessian matrix to penalize geometric deviations in complex regions.

Main Results:

• BA-UNet achieved a mean Dice Similarity Coefficient (DSC) of 85.41% and a 95% Hausdorff Distance (HD95) of 11.07 mm on an internal dataset.
• Demonstrated superior performance over state-of-the-art methods, especially for challenging anatomical structures.

Conclusions:

• The BA-UNet framework effectively improves segmentation accuracy for cervical cancer radiotherapy.
• The proposed BIFA module and BAC loss enhance geometric consistency, leading to better delineation of critical structures.