Deep learning-based computed tomography (CT) derived body composition classifier for colorectal cancer patients.

Accurate body composition analysis using computed tomography scans is critical for assessing nutritional status and predicting survival in cancer patients. Traditional body mass index measures are inadequate because they fail to distinguish between lean muscle mass and adipose tissue, which is particularly problematic in conditions like sarcopenic obesity where muscle depletion is masked by high body fat. While metrics such as skeletal muscle area and skeletal muscle density provide essential prognostic information, manual measurement is labor-intensive, requires specialist expertise, and is difficult to integrate into routine clinical workflows. This study explores the feasibility of using deep learning-based automated regression models to predict key body composition metrics from computed tomography scans of colorectal cancer patients.

The researchers evaluated four deep learning architectures—AlexNet, UNet, GoogLeNet, and ResNet34—to predict skeletal muscle area, skeletal muscle density, subcutaneous fat area, and visceral fat area. The dataset comprised 574 single-axial computed tomography scans at the third lumbar vertebra level from colorectal cancer patients treated at a single institution between 2008 and 2018. Standardized pre-processing, including Hounsfield Unit clipping and pixel normalization, ensured input consistency across scans. To perform numerical regression rather than classification or segmentation, the standard architectures were modified by replacing their final layers with a specialized regression head containing global average pooling and fully connected layers. Model optimization was systematically conducted using the Hyperband tuning algorithm alongside data augmentation strategies to improve generalization without introducing anatomical artifacts.

GoogLeNet achieved the strongest overall performance, recording the lowest mean percentage error for skeletal muscle area prediction at 4.96 percent. AlexNet performed best in predicting skeletal muscle density with a mean percentage error of 8.12 percent. Among all evaluated metrics, skeletal muscle area was predicted with the highest overall accuracy, whereas subcutaneous fat area presented the greatest challenge across all models. Following hyperband optimization and epoch adjustment, the best-performing models were integrated into a user-friendly web application built with Streamlit. Independent testing on a holdout set demonstrated an overall classification accuracy of 80% for categorizing body composition metrics into clinical risk scores.