LRF-UNet: Low-Rank Factorized Convolution Deep-Learning Networks for Visceral Adipose and Muscle Tissue Segmentation in Abdominal Computed Tomography Image.
Authors
Affiliations (4)
Affiliations (4)
- School of Medicine, Chung Shan Medical University, Taichung, 402, Taiwan.
- Department of Medical Imaging, Chung Shan Medical University Hospital, Taichung, 402, Taiwan.
- Department of Computer Science and Information Engineering, National Changhua University of Education, Changhua, 500207, Taiwan.
- Department of Mathematics, National Changhua University of Education, Changhua, 500207, Taiwan. [email protected].
Abstract
Quantitative analysis of skeletal muscle (SM) and visceral adipose tissue (VAT) cross-sectional volumes at the third lumbar vertebral level (L3) on abdominal computed tomography (CT) not only assists physicians in evaluating individual metabolic risk and nutritional status, but also aids clinicians in disease diagnosis, treatment planning, and prognostic evaluation. However, patient variability and inter-observer variability may lead to misdiagnosis. To enhance the robustness of quantitative analysis and clinical applicability, this study proposed a UNet-based deep learning segmentation system that automatically segments VAT and SM from single CT slices at the first through third lumbar vertebral levels (L1-L3). The proposed method consisted of image pre-processing, encoder, decoder, and the nested skip-connection module. In image pre-processing, the pixel intensities of input images were normalized through Z-score normalization, and data augmentation was applied before the training to increase data diversity. The encoder incorporated MobileNetV3 (MV3) blocks and introduced a low-rank factorized convolution (LRF-Conv) module to enhance feature representation of fragmented or indistinct tissue boundaries. Moreover, multi-scale feature fusion and gradient convergence were further improved through the LRF based decoder and nested skip-connection module, and the deep supervision strategy. The proposed method was trained and evaluated on the abdominal CT images from 179 patients, including 64 diabetic and 113 normoglycemic cases. To evaluate the generalizability of the model, fivefold cross-validation was performed in every experiment. The results yielded a Sørensen-Dice coefficient (Dice score) of 0.9435 ± 0.0045, an intersection over union (IoU) of 0.8973 ± 0.0056, and a 95th percentile Hausdorff distance (HD95) of 2.5325 ± 0.4702. These experimental results indicated that the proposed system revealed outstanding tissue segmentation performance in VAT and SM segmentation, thereby demonstrating potential for objective body composition analysis in routine clinical settings.