LRF-UNet: Low-Rank Factorized Convolution Deep-Learning Networks for Visceral Adipose and Muscle Tissue Segmentation...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology
• Body composition assessment

Background:

• Quantitative analysis of skeletal muscle (SM) and visceral adipose tissue (VAT) from abdominal CT scans is crucial for metabolic risk evaluation, nutritional status assessment, and clinical decision-making.
• Patient and inter-observer variability in manual segmentation can lead to diagnostic inaccuracies.

Purpose of the Study:

• To develop and evaluate a UNet-based deep learning system for automated segmentation of VAT and SM from CT slices at lumbar levels L1-L3.
• To enhance the robustness and clinical applicability of quantitative body composition analysis.

Main Methods:

• A UNet-based deep learning model incorporating MobileNetV3 (MV3) blocks and low-rank factorized convolution (LRF-Conv) was employed.
• Image pre-processing included Z-score normalization and data augmentation; the model featured an encoder-decoder structure with nested skip connections and deep supervision.
• The system was trained and validated on abdominal CT images from 179 patients (64 diabetic, 113 normoglycemic) using fivefold cross-validation.

Main Results:

• The system achieved a Sørensen-Dice coefficient (Dice score) of 0.9435 ± 0.0045 and an intersection over union (IoU) of 0.8973 ± 0.0056.
• A 95th percentile Hausdorff distance (HD95) of 2.5325 ± 0.4702 was recorded, indicating high segmentation accuracy.
• The deep learning system demonstrated outstanding performance in segmenting VAT and SM.

Conclusions:

• The proposed deep learning system offers robust and accurate automated segmentation of VAT and SM from abdominal CT images.
• This technology has significant potential for objective body composition analysis in routine clinical practice.
• The system can help overcome limitations associated with manual segmentation variability, improving diagnostic reliability.