Source-linked AI summary
Deep learning-based computed tomography (CT) derived body composition classifier for colorectal cancer patients
Eve Harling, Chattarin Pumtako, Bernd Porr, Donald C McMillan, Ross D Dolan
TL;DR
CT-based body composition analysis is clinically useful but traditionally labor-intensive and expertise-dependent. This pilot study evaluates deep learning models for automated prediction, with GoogLeNet achieving 4.96% percentage error for SMA prediction and 80% accuracy classifying low versus not-low SMI and SMD.
Problem
The full prognostic potential of CT-derived body composition remains limited in routine clinical assessment.
Method
The study applies deep learning-based CT analysis with standardized Hounsfield Unit clipping to improve muscle-tissue visibility.
Results
4.96% percentage error was achieved for SMA prediction by GoogLeNet, while classification of low versus not-low SMI and SMD reached 80% overall accuracy.
Takeaways & Limitations
GoogLeNet and AlexNet were identified as accurate models for automated body composition analysis.
Takeaways & Limitations
Performance across demographic subgroups requires further evaluation.
Abstract
from arXiv · showhide
Background: Accurate body composition analysis using Computed Tomography (CT) scans is essential for assessing skeletal muscle area (SMA) and skeletal muscle density (SMD), key markers of nutritional status in cancer patients. Conventional manual methods are labour-intensive and require specialist expertise, limiting their routine clinical use. Therefore, this study serves as a feasibility and pilot investigation to explore the potential of deep learning-based automated regression for body composition analysis within a clinical workflow. Methods: Four deep learning architectures (AlexNet, UNet, GoogLeNet, and ResNet34) were trained to predict SMA, SMD, subcutaneous fat area (SFA), and visceral fat area (VFA) from CT scans of colorectal cancer patients. Systematic hyperparameter optimization identified the most accurate models, which were subsequently implemented in a web application for clinical use. Results: GoogLeNet achieved the best performance, with a mean percentage error (PE) of 4.96% for SMA prediction, while AlexNet reached 8.12% for SMD. Independent testing demonstrated robust accuracy, correctly classifying body composition metrics in 80% of cases. The web application delivered rapid and consistent outputs, supporting integration into clinical workflows. Conclusion: Optimized deep learning models, particularly GoogLeNet and AlexNet, can automate CT-derived body composition analysis with a Mean Percentage Error (PE) of 4.96% for SMA and 8.12% for SMD. These tools have the potential to streamline clinical practice by reducing the time and expertise required for manual segmentation. Further validation in larger, more diverse datasets is warranted.
1. James Watt School of Engineering, College of Science & Engineering, University of Glasgow, Glasgow, UK
The section identifies a University of Glasgow academic affiliation in Glasgow, UK, alongside a Glasgow Royal Infirmary address and study keywords focused on deep-learning CT body-composition analysis.
- Affiliation: The Academic Unit of Surgery is part of the School of Medicine, College of Medical Veterinary & Life Sciences, University of Glasgow, Glasgow, UK.
- Address: The listed address is Room 2.51, New Lister Building, Glasgow Royal Infirmary, Glasgow, United Kingdom, G31 2ER.
- Keywords: Keywords include deep learning, computed tomography (CT), body composition analysis, colorectal cancer, and medical imaging.
Introduction
CT-derived body composition metrics provide clinically important information beyond BMI but remain underused because manual analysis is labour-intensive and requires specialized expertise. This study therefore evaluates deep learning architectures to develop a robust, clinically applicable tool for predicting body composition metrics from CT scans in colorectal cancer patients.
- Clinical barriers: CT-derived body composition analysis remains underutilized because it is time-consuming, labour-intensive, and dependent on specialized expertise and analytical resources.Limited access to advanced tools, high technology costs, and insufficient professional training further restrict routine adoption.
- Clinical importance: CT-derived SMA and SMD provide direct, precise quantification of physiological status, whereas BMI cannot distinguish lean muscle mass from adipose tissue.This limitation is particularly important in colorectal cancer, where sarcopenic obesity is associated with poor quality of life and reduced survival.
- Proposed solution: Deep learning, particularly convolutional neural networks, is proposed as a way to automate complex image analysis and incorporate detailed body composition assessment into routine clinical practice.These networks can learn to predict key body composition metrics from CT scans, potentially supporting cancer assessment and management.
- Study aim: The study evaluates AlexNet, UNet, GoogLeNet, and ResNet34 for predicting body composition metrics from CT scans of colorectal cancer patients.The stated goal is to develop a robust, clinically applicable tool that can enhance prognostic accuracy and improve oncology patient outcomes.
Methods
The study standardized and quality-controlled 574 L3 CT scans, then trained and optimized four CNN-based regression architectures to predict body-composition metrics. Patient-level splitting, controlled augmentation, and Hyperband tuning supported independent evaluation and clinical implementation.
- Model Selection and Evaluation: Four architectures—AlexNet, UNet, GoogLeNet, and ResNet34—were adapted with regression heads to predict continuous SFA, VFA, SMA, and SMD values.Models were evaluated using Mean Squared Error (MSE) and Percentage Error (PE).
- Model Selection and Evaluation: 80/20 training-validation splitting was performed at the patient level, with 10 patients held out for independent prediction testing.All scans from each patient were assigned exclusively to one dataset partition to prevent data leakage.
- Model Optimization: Hyperband systematically explored filter numbers, dropout rates, and learning rates to identify configurations minimizing validation loss under computational constraints.Initial settings included batch size 16, learning rate 0.0001, and 100 epochs.
- Data Augmentation and Training: Controlled augmentation used horizontal flips, rotations within ±10 degrees, and slight zooms to improve generalization while preserving anatomical plausibility.The models were trained with MSE recorded at each epoch as the loss function.
Results
GoogLeNet and AlexNet achieved the strongest body-composition prediction performance, with GoogLeNet leading overall and optimized models supporting CT-SS classification with 80% accuracy. Performance varied by metric, subgroup, and computational cost, while independent testing showed strong agreement for SMI and SMD predictions.
- Model Performance: 4.96% PE was achieved by GoogLeNet for SMA prediction (MSE: 450), while its SMD prediction reached 8.04% PE (MSE: 34).AlexNet achieved 8.12% PE for SMD (MSE: 40), whereas ResNet34 had 12.02% PE for VFA (MSE: 5935).
- CT-SS Classifier Results: 80% overall accuracy was achieved by the CT-SS classifier, with high sensitivity for Score 2 and most false positives occurring in Score 1.The balanced F1-score was 0.75, while errors were more frequent among female patients for SMI and male patients with BMI < 25 for SMD.
Discussion
GoogLeNet was most suitable for SMA prediction, while AlexNet performed best for SMD, with PE values of 4.96% and 8.12%, respectively. The study also demonstrated rapid clinical outputs while identifying limitations requiring broader validation and refinement.
- Model performance: GoogLeNet was most suitable for SMA prediction, whereas AlexNet was the top performer for SMD.These selections were based on comparative model performance.
- Prediction accuracy: 4.96% PE was achieved by GoogLeNet for SMA, while AlexNet achieved 8.12% PE for SMD.SFA had the least accurate predictions, with a mean PE of 10.9%.
- Clinical application: Predictions were generated in under a second per scan, and CT-SS predictions matched manually calculated values.The application was described as computationally efficient, intuitive, and promising for clinical use, but dependent on correct input formatting and lacking safeguards against erroneous inputs.
- Limitations and future work: The single-centre Glasgow Royal Infirmary dataset and small web-application evaluation limit generalizability and statistical power.Future work will prioritize multi-centre validation, broader demographics, formal agreement metrics, and improved SFA and VFA accuracy.
Ethics Statement
The study received ethics approval and complied with the Declaration of Helsinki, with informed consent waived for retrospective analysis of anonymized routine clinical data. Data and source code access are specified, and the authors report no conflicts or specific funding.
- The West of Scotland Research Ethics Committee, Glasgow, reviewed and approved the study, which was conducted in accordance with the Declaration of Helsinki.
- Written informed consent was waived under national legislation and committee requirements because the study retrospectively analyzed anonymized routine clinical data.
- Datasets are available from the corresponding author on reasonable request, while analysis source code is publicly available at Zenodo.
- The authors declare no conflict of interest and received no specific funding for this work.