A multitask deep learning framework for the automated quantification of abdominal tissue morphology on ultrasound and the therapeutic monitoring of diastasis recti abdominis.
Authors
Affiliations (6)
Affiliations (6)
- Department of Ultrasound, Huadu District People's Hospital of Guangzhou, Guangzhou, China.
- Huadu Institute of Medical Sciences, Guangzhou, China.
- Department of Plastic and Reconstructive Surgery, The Affiliated Guangdong Second Provincial General Hospital of Jinan University, Guangzhou, China.
- The Chinese University of Hong Kong (Shenzhen), Shenzhen, China.
- Shenzhen Yingzhichuangsi Technology Co., Ltd., Shenzhen, China.
- Department of Radiology, The First Affiliated Hospital of Guangzhou Medical University, Guangzhou, China.
Abstract
Diastasis recti abdominis (DRA) requires accurate ultrasound-based quantification of abdominal tissue morphology, but manual measurement is operator-dependent and inefficient for longitudinal monitoring. This study developed and evaluated a multitask deep learning model for automated abdominal ultrasound analysis and the monitoring of magnetic stimulation therapy. A retrospective dataset of 2,750 abdominal ultrasound images from 632 women was used to train a ResNet-50-encoded U-Net with dual branches for tissue segmentation (linea alba, rectus abdominis, and subcutaneous fat) and anatomical region classification. Model performance was evaluated on an independent set of 550 images. The trained model was then applied prospectively to 21 patients undergoing six sessions of targeted abdominal magnetic stimulation therapy, with ultrasound measurements obtained before treatment, 1 day after the first session, and 1 day after the sixth session. The model achieved Dice coefficients of 0.979±0.013 for subcutaneous fat, 0.951±0.066 for rectus abdominis muscle, and 0.828±0.074 for the linea alba. Automated measurements showed strong agreement with expert measurements, with intraclass correlation coefficients (ICCs) of 0.988 for subcutaneous fat thickness (SET), 0.955 for the interrectus distance (IRD), and 0.917 for rectus abdominis muscle thickness (RAMT). After six therapy sessions, the IRD decreased from 20.17±7.81 to 15.22±6.58 mm (P=0.04). Left and right RAMT increased from 8.27±1.04 to 10.26±1.23 mm and from 8.22±1.11 to 10.19±1.45 mm, respectively (both P values <0.001). There was a nonsignificant (P=0.05) decrease in SET from 16.56±3.56 to 14.30±3.55 mm. The proposed multitask model enabled automated multitissue quantification on abdominal ultrasound images with strong agreement with expert measurements. It may support reproducible diagnosis and longitudinal monitoring of patients with DRA. Meanwhile, its ability to monitor treatment-related changes, particularly increases in short-term muscle thickness and reductions in subcutaneous fat, should be interpreted cautiously and validated in larger controlled studies.