A Lightweight Skeletal Muscle Intelligent Segmentation Network Based on Planning CT for Cervical Cancer Radiotherapy.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiotherapy Planning

Background:

• Skeletal muscle segmentation is crucial for assessing sarcopenia in cancer patients.
• Accurate L3 skeletal muscle area (SMA) measurement aids in treatment planning and outcome prediction.
• Current segmentation methods can be time-consuming and require significant expertise.

Purpose of the Study:

• To develop and evaluate a lightweight deep learning network, SMA-Net, for automated L3 skeletal muscle segmentation.
• To assess the segmentation performance and diagnostic utility of SMA-Net in cervical cancer patients.
• To compare SMA-Net's efficiency and accuracy against manual segmentation and existing networks.

Main Methods:

• A lightweight Mamba architecture integrated into a UNet network (SMA-Net) was developed.
• SAB and CAB attention mechanisms were incorporated to enhance feature discrimination.
• The network was trained and validated on 160 cervical cancer patients' radiotherapy images.
• Segmentation performance was evaluated using Dice similarity coefficient, sensitivity, PPV, and Hausdorff distance.

Main Results:

• SMA-Net achieved a Dice similarity coefficient of 89.16% for skeletal muscle segmentation.
• The network demonstrated high accuracy (87.5%) in predicting sarcopenia in cervical cancer patients.
• SMA-Net requires only 1.50 GFLOPS and 1.24 M parameters, indicating computational efficiency.
• Radiologist review showed 93.75% of segmentations required minor or no revisions.

Conclusions:

• The lightweight SMA-Net accurately and efficiently segments L3 skeletal muscle, facilitating clinical application.
• SMA-Net aids in the rapid diagnosis of sarcopenia in cervical cancer patients.
• This tool can potentially reduce physician workload and improve diagnostic efficiency in clinical settings.