Development of a fully automatic deep learning system for L3 selection and body composition assessment on computed...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Sarcopenia research increasingly requires abdominal muscle quantification using computed tomography (CT).
• Automated systems for L3 slice selection and muscle segmentation are needed to address this demand.
• Existing methods may not adequately account for anatomical variations.

Purpose of the Study:

• To develop a deep learning model (DLM) for automated L3 slice selection and abdominal muscle/fat segmentation on CT.
• To evaluate the DLM's accuracy, considering anatomical variations.
• To provide an end-to-end automated solution for sarcopenia-related imaging analysis.

Main Methods:

• Developed L3SEG-net, a DLM combining YOLOv3 for L3 slice selection and a fully convolutional network (FCN) for segmentation.
• Trained the YOLOv3 component on a dataset of 922 CT scans.
• Validated the model on internal (496 scans) and external (586 scans) datasets, with radiologist-identified ground truths.

Main Results:

• High accuracy in L3 slice selection (mean distance difference < 5mm) and technical success rates >92% in validation datasets.
• Reduced accuracy in L3 slice selection (mean distance difference >12mm) and technical success rates ~67% in subgroups with anatomical variations.
• Excellent segmentation accuracy for abdominal muscle areas, with low cross-sectional area (CSA) errors (1.38-3.10 cm²), irrespective of anatomical variations.

Conclusions:

• The developed L3SEG-net system automates L3 slice selection and abdominal muscle segmentation on CT scans.
• The system demonstrates high accuracy, though anatomical variations present a challenge for precise L3 slice localization.
• The robust muscle segmentation performance suggests utility in sarcopenia research and clinical applications.