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

Deep learning-based computed tomography (CT) derived body composition classifier for colorectal cancer patients | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Deep learning-based computed tomography (CT) derived body composition classifier for colorectal cancer patients Eve Harling, Chattarin Pumtako, Bernd Porr, Donald C McMillan, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7585523/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 13 You are reading this latest preprint version Abstract 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. 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 high accuracy. 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. Deep learning Computed tomography (CT) Body composition analysis Colorectal cancer Medical imaging Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Computed Tomography (CT) scans have emerged as invaluable tools in cancer management, providing detailed insights into patient anatomy that are crucial for prognosis and treatment planning. However, the full prognostic potential of CT-derived body composition analysis remains underutilized in clinical practice. This underutilization is largely due to the time-consuming nature of the analysis and the specialized expertise required to accurately interpret the data ( 1 ). Furthermore, the lack of access to advanced analytical tools, the high costs associated with these technologies, and insufficient training among healthcare professionals further limit the widespread adoption of these critical assessments ( 2 ). Historically, Body Mass Index (BMI) has been the standard metric used for assessing the nutritional and physical status of cancer patients. However, with the increasing prevalence of obesity, BMI has been shown to be an inadequate measure for predicting outcomes in oncology. Unlike BMI, CT-derived metrics provide a more nuanced and clinically relevant assessment of a patient’s body composition ( 3 ). These metrics include skeletal muscle area (SMA), skeletal muscle density (SMD), subcutaneous fat area (SFA), and visceral fat area (VFA), all of which offer critical insights into a patient’s health status. The importance of these metrics extends beyond simple body composition. For example, sarcopenia, which is characterized by the degenerative loss of muscle mass and function, is a common condition in cancer patients that is associated with poor clinical outcomes, including reduced survival rates ( 4 , 5 ). Similarly, cachexia, which involves a disproportionate loss of skeletal muscle relative to fat, significantly impacts a patient’s quality of life and overall prognosis ( 6 ). CT-derived metrics such as SMD can also indicate myosteatosis, a condition where muscle tissue is infiltrated by fat, further complicating a patient's clinical outlook. One of the most significant CT-derived metrics is the Skeletal Muscle Index (SMI), which normalizes the SMA to a patient's height squared, providing a standardized measure of muscle mass that is strongly associated with survival in cancer patients. When combined with SFA, VFA, and SMD, these metrics can be used to calculate a CT-derived sarcopenia score (CT-SS), which serves as a powerful prognostic tool. The CT-SS is categorized into three levels, with higher scores indicating poorer prognosis. Despite their proven utility, the application of these measures in routine clinical practice remains limited due to the labour-intensive nature of manual analysis and the complexity of integrating these metrics into clinical workflows ( 7 ). The advent of deep learning technologies offers a promising solution to these challenges. Deep learning, particularly convolutional neural networks (CNNs), has demonstrated remarkable success in automating complex image analysis tasks, making it feasible to incorporate detailed body composition assessments into routine clinical practice. By leveraging large datasets, these networks can learn to accurately predict key body composition metrics from CT scans, potentially transforming the way clinicians assess and manage cancer patients ( 8 ). This study aims to evaluate the effectiveness of several deep learning architectures, including AlexNet ( 9 ), UNet ( 10 ), GoogLeNet ( 11 ), and ResNet34 ( 12 ) in predicting body composition metrics from CT scans of colorectal cancer patients. The ultimate goal is to develop a robust, clinically applicable tool that can enhance prognostic accuracy and improve patient outcomes in oncology. Methods Data Preparation : This study utilized two primary datasets: one consisting of CT scans from the third lumbar vertebra (L3) of patients diagnosed with colorectal cancer at the Glasgow Royal Infirmary between 2008 and 2018, and another comprising anonymized patient information in a tabulated format. Initially, the CT scans were in DICOM format, which was incompatible with TensorFlow, the deep learning framework used in this study. To address this, the images were batch-converted to JPEG format using ImageJ ( 13 ). To optimize computational efficiency, the images were resized from 512×512 pixels to 256×256 pixels. A total of 574 CT scans were retained after processing, followed by a visual inspection for anisometry errors. Ten scans were randomly selected and set aside as a prediction set for post-training model evaluation. Model Selection and Evaluation Four well-established deep learning architectures were selected for this study. AlexNet ( 9 ) was chosen as a benchmark due to its simplicity and efficiency, achieved through the use of convolutional layers, ReLU activations, and Max-Pooling layers. These features make AlexNet particularly well-suited for handling large-scale image data with relatively low computational costs. UNet ( 10 ) was included for its strength in biomedical image segmentation, especially its ability to accurately extract detailed features from images, which is critical for segmenting body composition metrics like SFA and VFA from CT scans. GoogLeNet ( 11 ) was selected for its innovative inception modules, which enable the model to capture multi-scale features within a single layer. This capability makes it particularly effective in distinguishing between different tissue types in complex CT images, while its use of global average pooling helps prevent overfitting. ResNet34 ( 12 ) was chosen for its depth and the use of residual connections, which help mitigate the vanishing gradient problem in deep networks, making it suitable for capturing intricate patterns in CT scan data, particularly for SMD and SMA predictions. The models were evaluated based on their performance in predicting SFA, VFA, SMA, and SMD, using Mean Squared Error (MSE) and Percentage Error (PE) as metrics. Given that these output variables are continuous, regression models were constructed for each architecture. The training process commenced with the setup of TensorFlow and Keras, the core libraries used for developing and fine-tuning the deep learning models. The CT scan data were carefully pre-processed and organized into a structured dataset, which included converting DICOM images to JPEG format and resizing them to 256×256 pixels to optimize for computational efficiency. Hyperparameters, including batch size of 16 ( 14 ), learning rate (0.0001), and epochs (100), were uniformly initialized across all models. A seed value of 42 was used to ensure reproducibility. Data Augmentation and Training To improve the models' ability to generalize, data augmentation techniques such as random flips, rotations, and zooms were applied. These augmentations aimed to reduce overfitting by introducing variability into the training data. The models were built using the Keras functional API, which provides flexibility in model architecture design. An 80/20 data split was used for training and validation, with the MSE recorded at each epoch as the loss function. Model Optimization Following the initial training, GoogLeNet demonstrated the highest accuracy in predicting SMA, making it the focus of further optimization using Hyperband tuning ( 15 ). Hyperband was selected for its efficiency in optimizing hyperparameters under limited computational resources. The key hyperparameters optimized for GoogLeNet included filter numbers, learning rate, and seed. Similarly, AlexNet, which performed well in predicting SMD, underwent optimization focusing on dropout rate, dense layer nodes, and learning rate. Development of the CT-SS Classifier Post-optimization, the most accurate models—GoogLeNet for SMA and AlexNet for SMD—were used to develop a CT-SS classifier. The CT-SS was calculated using the SMI, derived by normalizing SMA by the square of the patient’s height. SMI and SMD values were classified as low or not low based on established thresholds dependent on BMI, sex, and height. The final CT-SS score, ranging from 0 to 2, was computed, with scores of 2 indicating low SMI and SMD, 1 indicating low SMI or SMD, and 0 indicating normal or high values in both metrics. Web Application Development To enable the practical application of the developed models in clinical settings, a web-based application was created using Streamlit. This app allows clinicians to input patient data, including sex, height, BMI, and upload CT scans for analysis. The app predicts SMA and SMD using the trained models, calculates SMI, and provides a CT-SS score. The user-friendly interface ensures that the tool can be seamlessly integrated into clinical practice, facilitating quick and accurate assessments. Results Model Performance : The performance of the four deep learning models was evaluated based on their ability to predict SFA, VFA, SMA, SMD from CT scans. The evaluation metrics used included MSE and PE. AlexNet excelled in predicting SMD, achieving a PE of 8.12% and an MSE of 40. For SMA, AlexNet recorded a PE of 6.15% and an MSE of 468. UNet showed competitive performance, particularly in predicting SFA and VFA, with a PE of 8.76% (MSE: 3384) for SFA and a PE of 9.52% (MSE: 3998) for VFA. GoogLeNet demonstrated the strongest overall performance, achieving the lowest PE across multiple metrics. Specifically, GoogLeNet predicted SMA with a PE of 4.96% and an MSE of 450. For SMD, it achieved a PE of 8.04% and an MSE of 34. ResNet34 recorded a PE of 10.30% for SFA (MSE: 3493) and a PE of 10.56% for SMD (MSE: 40). It particularly struggled with predicting VFA, where it had a PE of 12.02% and an MSE of 5935. Table 1 summarizes these results. Figure 1 a illustrates the PE for each model across all body composition metrics, highlighting that GoogLeNet achieved the lowest overall PE at 8.04%, followed closely by AlexNet at 8.52%. In contrast, UNet and ResNet34 had higher PEs at 10.30% and 11.56%, respectively, while Fig. 1 b illustrates the mean PE for each body composition metric. The architectures encountered the greatest difficulty with SFA, with a mean PE of 10.9%. The best performance was with SMA, making it the most accurate predicted overall with a mean PE of 7.29%. Table 1 Comparison of the Different Models AlexNet GoogLeNet UNet ResNet34 MSE AE PE% MSE AE PE% MSE AE PE% MSE AE PE% SFA (cm²) 3412 58.41 8.16 6869 82.88 11.58 5935 77.04 10.76 9896 99.48 13.90 VFA (cm²) 3998 63.23 9.52 1498 38.70 5.83 3493 59.10 8.90 3384 58.17 8.76 SMA (cm²) 468 21.63 6.27 450 21.21 6.15 882 29.70 8.60 868 29.46 8.53 SMD (HU) 34 5.83 9.28 40 6.32 10.07 57 7.55 12.02 44 6.63 10.56 Time Efficiency The computational experiments were performed using an Intel(R) Xeon(R) CPU E5630 @ 2.53GHz and a GeForce GTX 1070 GPU. The combination of these hardware components allowed for efficient processing of the deep learning models. The time required per epoch for each model was recorded to assess computational efficiency. AlexNet completed each epoch in 1 second, making it the fastest model, GoogLeNet completed each epoch in 3 second, UNet completed each epoch in 12 second, whereas ResNet34 required 60 seconds per epoch, making it the slowest. Hyperparameter Optimization Hyperband hyperparameter tuning was applied to optimize the GoogLeNet model for SMA and the AlexNet model for SMD. After optimization, GoogLeNet's SMA prediction error was reduced from 6.15% to 5.65%, and AlexNet's SMD prediction error decreased from 9.28% to 8.96%. Model Customization and Final Performance Further customization of the models was performed to enhance their performance. Figure 2 a illustrates the change in PE when increasing the number of epochs for GoogLeNet with SMA. Although 1000 epochs exhibited the lowest PE, it demonstrated a high degree of unreliability and experienced interruptions during training due to out of memory (OOM) errors. A local minimum is found at 300 epochs that was consistent across multiple reruns and had no reliability issues. Given the selection of 300 epochs for GoogLeNet. Figure 2 b shows the change in PE for AlexNet training with SMD at different epochs. The lowest PE value, 8.12%, achieved was at 600 epochs. there were no OOM errors encountered during training. Given the selection of 600 epochs for AlexNet, these values were determined through iterative testing and careful analysis of the models' learning curves. Table 2 summarises the overall impact of these changes on the performance of the GoogLeNet and AlexNet models. Table 2 Impact of Changes implemented Change Made GoogLeNet SMA Model AlexNet SMD Model PE% Achieved Improvement % PE% Achieved Improvement % Original 6.15 - 9.28 - HyperBand Optimizing 5.65 -0.50 8.96 -0.32 Alternative End 5.65 0 8.28 -0.68 Number of Epochs 4.96 -0.69 8.12 -0.16 Batch Size 4.96 0 8.61 0 Total 1.19 1.16 CT-SS Classifier Results The CT-SS classifier, developed using the optimized GoogLeNet and AlexNet models, was tested on a separate set of CT scans. The mean absolute error for SMI predictions was 2.299, with a relative mean error of 5.02% Fig. 3 a. The regression equation indicates that the associated p-value = 0.0005 for SMI predictions and true SMI Fig. 4 a. For SMD predictions the mean absolute error was 1.852, resulting in a relative mean error of 9.85% Fig. 3 b. The regression equation indicates that the associated p-value = 0.0004 for SMD predictions and true SMD Fig. 4 b. The classifier achieved an overall accuracy of 80% for the CT-SS classifications. Web Application Deployment A Model application was developed to integrate the trained models into a user-friendly interface for clinical use. The application allows users to input patient data, including sex, height, and BMI, and upload CT scans to receive predictions for SMA, SMD, and the CT-SS score. Discussion The deep learning models employed in this study were selected based on their balance between accuracy and computational efficiency, which are critical factors for practical clinical application. Among the architectures tested, GoogLeNet emerged as the most suitable for predicting SMA, while AlexNet was the top performer for SMD. These selections were driven by their performance metrics and computational demands during training. GoogLeNet's superior performance can be attributed to its innovative use of parallel layers with multiple filter sizes, which allows the model to process input data at various scales within a single layer. This design reduces the computational burden by incorporating global average pooling layers, which minimize the number of trainable parameters and help prevent overfitting. Additionally, the reliance on small, computationally inexpensive convolutions further enhances GoogLeNet's efficiency, making it comparable to AlexNet despite its deeper architecture. AlexNet, although simpler and shallower than the other models, demonstrated the highest accuracy in predicting SMD. This is likely due to its compact architecture, which reduces the overall number of trainable parameters and accelerates the optimization process. The use of Max-Pooling and dropout layers further contributes to its computational efficiency by limiting the number of nodes and computations required per epoch. Interestingly, the relatively shallow depth of AlexNet may have allowed it to avoid overfitting on non-relevant features, thus maintaining high accuracy in SMD prediction. Conversely, ResNet34, the deepest model tested, underperformed both in terms of accuracy and computational efficiency. The absence of layers designed to reduce computational costs, such as Max-Pooling or Global Average Pooling (GAP), combined with its substantial depth, likely contributed to its lower efficiency and accuracy. This underperformance suggests that ResNet34 may have encountered degradation issues, where the added depth did not translate into improved accuracy, potentially due to noise or irrelevant patterns in the data being learned by the model. The prediction accuracy varied across the body composition metrics, with SMA and SMD being predicted more accurately than SFA and VFA. The PE for SMA was notably lower, especially with the GoogLeNet model, which achieved a PE of 6.15%. This performance underscores the model's robustness in capturing muscle-related features from CT scans. In contrast, SFA predictions were the least accurate, with a mean PE of 10.9%, proving particularly challenging for both GoogLeNet and ResNet34. The difficulty in distinguishing between SFA and VFA, given their similar appearances on CT scans, may have contributed to this higher error rate. Additionally, the edge detection required to separate SFA from the background might have been a limiting factor for these models. Despite these challenges, VFA was predicted with slightly better accuracy, indicating that the models could more reliably detect fat distributions closer to the visceral organs. SMD predictions, while more accurate than SFA, still presented challenges, particularly due to the nature of the metric, which relies on the brightness of the muscle in Hounsfield units. The percentage error for SMD averaged 10.44%, with AlexNet achieving the lowest error at 9.28%. This suggests that while the models are proficient at detecting muscle mass, further refinement is necessary to accurately capture muscle density. The Hyperband optimization method proved effective in enhancing model performance, particularly for the GoogLeNet SMA model, where the percentage error was reduced by 0.459%. This improvement was achieved by fine-tuning the initial and final filter numbers, which increased the number of trainable parameters, leading to better accuracy without compromising computational efficiency. The learning rate, a critical parameter for balancing convergence speed and precision, was also optimized, contributing to the model's improved performance. For the AlexNet SMD model, the Hyperband tuning resulted in a smaller but significant improvement of 0.32%. The optimization focused on the dense and dropout layers, with the most substantial gain in accuracy (0.68%) achieved by replacing the fully connected, dense layers with those from GoogLeNet. This substitution not only enhanced accuracy but also reduced computational costs, demonstrating the effectiveness of combining elements from different architectures to leverage their strengths. The most considerable improvement for the GoogLeNet SMA model came from increasing the number of training epochs, which reduced the percentage error by 0.69%. However, the relationship between epoch count and error reduction was non-linear, with diminishing returns observed beyond certain thresholds, likely due to overfitting or noise in the dataset being misinterpreted as relevant features. Similar trends were observed for AlexNet, where the percentage error decreased up to 600 epochs but increased slightly thereafter, likely due to overfitting. The CT-SS classifier, developed using the optimized GoogLeNet and AlexNet models, demonstrated an overall accuracy of 80% in classifying SMI and SMD as "low" or "not low." The classifier's performance, as indicated by the F1 score of 0.75, was reasonably balanced, though there is room for improvement, particularly in reducing false positives. The errors observed in classification were concentrated in specific patient subgroups, notably female patients for SMI and male patients with a BMI less than 25 for SMD. These findings suggest that further research is needed to explore sex- and BMI-related variations in model accuracy, potentially requiring tailored approaches for different demographic groups. The developed web application served as a proof of concept for integrating deep learning models into a user-friendly clinical tool. The app's performance was consistent, with predictions generated in under a second per scan, and the CT-SS predictions matched those calculated manually from the predicted values. However, the app's reliance on correct input data format and the lack of safeguards against erroneous inputs are notable limitations. Despite these issues, the app's intuitive interface and robust performance in handling valid data make it a promising tool for clinical use, provided further refinements are made to address its current limitations. In comparing our results with existing work in the field, we found that several studies have applied deep learning models to body composition analysis, particularly using abdominal CT scans. For instance, Hsu et al. (2021) developed a machine learning model to predict body composition metrics such as visceral fat area (VFA), subcutaneous fat area (SFA), and skeletal muscle area (SMA) from CT scans of patients with pancreatic cancer. However, their approach differs from ours in several key aspects: their study utilized a U-Net architecture to perform image segmentation ( 16 ), whereas our work adopts a black-box approach, providing a numerical output for body composition metrics, including skeletal muscle density (SMD). Additionally, our approach achieved a superior performance, particularly for SMA prediction, where GoogLeNet attained a percentage error (PE) of 6.15%. Similarly, Dabiri et al. (2019) applied machine learning techniques to segment muscles and fat tissues from CT scans at the lumbar (L3) and thoracic (T4) levels in patients with colorectal cancer. Their approach also relied on image segmentation, focusing on muscle segmentation at L3, like our study ( 17 ). However, our work distinguishes itself by focusing on both segmentation and the prediction of numerical metrics, such as SMD and fat area, with AlexNet showing the highest accuracy for SMD prediction, achieving a PE of 9.28%. Although Dabiri et al. did not provide detailed performance metrics comparable to ours, it is clear that our black-box numerical prediction approach offers a novel contribution, particularly for SMD prediction, which remains an underexplored area. while previous work has primarily focused on image segmentation for body composition analysis, our study presents a novel approach by providing numerical outputs for metrics such as SMA, SFA, VFA, and SMD. Our results indicate that the deep learning models tested, particularly GoogLeNet and AlexNet, offer competitive and in some cases superior accuracy to existing segmentation-based approaches, particularly in the prediction of SMA and SMD. This highlights the practical applicability of our approach in clinical settings, where rapid and accurate predictions of body composition metrics are essential for assessing patient health, especially in individuals with conditions like colorectal cancer. In conclusion, the results of this study demonstrate the feasibility of using deep learning models, specifically GoogLeNet and AlexNet, for accurate body composition analysis in clinical settings. The optimization processes applied have significantly enhanced the models' performance, making them more suitable for practical application. The study underscores the potential of these deep learning models to automate the analysis of CT-derived body composition metrics, which are crucial for assessing patient health, particularly in individuals with colorectal cancer. By accurately predicting these metrics, the models can assist clinicians in making informed decisions regarding patient prognosis, treatment planning, and monitoring the progression of conditions such as sarcopenia and cachexia. The models' ability to efficiently handle large-scale image data positions them as viable candidates for integration into clinical workflows. However, the study also highlights several areas for future research. There remains room for further optimization, particularly in improving the prediction accuracy of SFA and VFA, where the models exhibited higher PEs. Future work could explore the use of more advanced data augmentation techniques or the incorporation of transfer learning to further enhance model performance. Additionally, expanding the dataset to include a more diverse range of patient demographics and health conditions could improve the models' generalization capabilities. Moreover, real-world clinical validation is essential to ensure their robustness and reliability in everyday practice. The development of a user-friendly web application, as outlined in the study, represents a significant step toward this goal by providing clinicians with a practical tool for quick and accurate body composition analysis. Declarations Ethics Statement This study was reviewed and approved by the West of Scotland Research Ethics Committee, Glasgow. All research was conducted in accordance with the Declaration of Helsinki. Written informed consent for participation was waived in accordance with national legislation and the requirements of the West of Scotland Research Ethics Committee because the study involved retrospective analysis of anonymized data collected during routine clinical care. Consent for publication : Not applicable. Competing Interests: The author(s) declare no conflict of interest. Funding: The author(s) received no specific funding for this work. Author Contribution EH designed the work, acquired data, interpreting the results, revised the manuscript, Approved the final version. CP interpreting the results, drafted or revised the manuscript, Approved the final version. BP designed the work, revised the manuscript, Approved the final version. DCM designed the work, acquired data, interpreting the results, revised the manuscript, Approved the final version. RDD designed the work, acquired data, interpreting the results, revised the manuscript, Approved the final version. Acknowledgements: The authors have nothing to report. Data Availability The datasets generated and/or analysed during the current study are available from the corresponding author on reasonable request. The source code used for the analysis is publicly available at Zenodo ( [https://zenodo.org/records/10439107](https:/zenodo.org/records/10439107) ) (18) References McGovern J, Delaney J, Forshaw MJ, McCabe G, Crumley AB, McIntosh D, et al. The relationship between computed tomography-derived sarcopenia, cardiopulmonary exercise testing performance, systemic inflammation, and survival in good performance status patients with oesophago-gastric cancer undergoing neoadjuvant treatment. JCSM Clin Rep. 2023;8(1):3–11. McGovern J, Golder AM, Dolan RD, Roxburgh CSD, Horgan PG, McMillan DC. The combination of computed tomography-derived muscle mass and muscle density and relationship with clinicopathological characteristics and survival in patients undergoing potentially curative surgery for colorectal cancer. JCSM Clin Rep. 2022;7(3):65–76. Arends J, Baracos V, Bertz H, Bozzetti F, Calder PC, Deutz NEP, et al. ESPEN expert group recommendations for action against cancer-related malnutrition. Clin Nutr. 2017;36(5):1187–96. Tolonen A, Pakarinen T, Sassi A, Kytta J, Cancino W, Rinta-Kiikka I, et al. Methodology, clinical applications, and future directions of body composition analysis using computed tomography (CT) images: A review. Eur J Radiol. 2021;145:109943. Vergara-Fernandez O, Trejo-Avila M, Salgado-Nesme N. Sarcopenia in patients with colorectal cancer: A comprehensive review. World J Clin Cases. 2020;8(7):1188–202. Fearon K, Strasser F, Anker SD, Bosaeus I, Bruera E, Fainsinger RL, et al. Definition and classification of cancer cachexia: an international consensus. Lancet Oncol. 2011;12(5):489–95. Somasundaram E, Castiglione JA, Brady SL, Trout AT. Defining Normal Ranges of Skeletal Muscle Area and Skeletal Muscle Index in Children on CT Using an Automated Deep Learning Pipeline: Implications for Sarcopenia Diagnosis. AJR Am J Roentgenol. 2022;219(2):326–36. Dolan RD, Almasaudi AS, Dieu LB, Horgan PG, McSorley ST, McMillan DC. The relationship between computed tomography-derived body composition, systemic inflammatory response, and survival in patients undergoing surgery for colorectal cancer. J Cachexia Sarcopenia Muscle. 2019;10(1):111–22. Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Commun ACM. 2017;60(6):84–90. Ronneberger O, Fischer P, Brox T, editors. U-Net: Convolutional Networks for Biomedical Image Segmentation. Cham: Springer International Publishing; 2015. Szegedy C, Liu W, Jia Y, Sermanet P, Reed SE, Anguelov D et al. Going deeper with convolutions. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2014:1–9. He K, Zhang X, Ren S, Sun J. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2015:770-8. Schneider CA, Rasband WS, Eliceiri KW. NIH Image to ImageJ: 25 years of image analysis. Nat Methods. 2012;9(7):671–5. Kingma DP, Ba J, Adam. A Method for Stochastic Optimization. CoRR. 2014;abs/1412.6980. Li L, Jamieson KG, DeSalvo G, Rostamizadeh A, Talwalkar A. Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization. J Mach Learn Res. 2016;18:1851. Hsu TH, Schawkat K, Berkowitz SJ, Wei JL, Makoyeva A, Legare K, et al. Artificial intelligence to assess body composition on routine abdominal CT scans and predict mortality in pancreatic cancer- A recipe for your local application. Eur J Radiol. 2021;142:109834. Dabiri S, Popuri K, Cespedes Feliciano EM, Caan BJ, Baracos VE, Beg MF. Muscle segmentation in axial computed tomography (CT) images at the lumbar (L3) and thoracic (T4) levels for body composition analysis. Comput Med Imaging Graph. 2019;75:47–55. Harling E. 2451792h_MastersProject [Model]. Zenodo: University of Glasgow; 2023. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 12 Feb, 2026 Reviewers agreed at journal 12 Feb, 2026 Reviews received at journal 11 Feb, 2026 Reviewers agreed at journal 01 Feb, 2026 Reviews received at journal 12 Jan, 2026 Reviewers agreed at journal 12 Jan, 2026 Reviewers agreed at journal 10 Dec, 2025 Reviewers agreed at journal 09 Nov, 2025 Reviewers invited by journal 13 Oct, 2025 Editor assigned by journal 08 Oct, 2025 Editor invited by journal 23 Sep, 2025 Submission checks completed at journal 22 Sep, 2025 First submitted to journal 16 Sep, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7585523","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":535112716,"identity":"84b4a6a8-4726-4a3f-a241-2a49deded4e0","order_by":0,"name":"Eve Harling","email":"","orcid":"","institution":"University of Glasgow","correspondingAuthor":false,"prefix":"","firstName":"Eve","middleName":"","lastName":"Harling","suffix":""},{"id":535112717,"identity":"fd512e28-953d-48a5-8495-dfeed52b0237","order_by":1,"name":"Chattarin Pumtako","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABMElEQVRIie3RsUrDQBjA8TsOkuXg1i8K7StcCUTBqH2UhECmLG6FSnohkC7FPoCCz+AbRAJ1OZwtLumSqUO7FSlirgYRkoBuIvlDyBG+H3dHEOrq+oMxwXL15gjhKK8+auVzWq3TGoGUltOOIiTm3wi0Ej78InoCPyOIWkfL3YXJEE7Go1HYY7dZkaM36CM9ywmVDYQVx4HjWYbAyVLKzIRn/4TjGxgI6nNCX5p20UpCbJ6WJEpSV0iqAZ4BFihAhG7ayORArqL3cHJfkUvB1s1keCCZpQiOBHG4ImgHrgC1S/1gkFFyFvhPphHj2BCLbPAgNUvNewkU/PGufn02neHXwB4P5vp0tRXXYb8nSQGbvX0+Z94qXy9qBJHaQuUmn38nrYPW9r+Y7erq6vrvfQDvlWG3PMeBowAAAABJRU5ErkJggg==","orcid":"","institution":"University of Glasgow","correspondingAuthor":true,"prefix":"","firstName":"Chattarin","middleName":"","lastName":"Pumtako","suffix":""},{"id":535112718,"identity":"35c30d9a-2a14-40ce-bcc0-83bf5cf5761c","order_by":2,"name":"Bernd Porr","email":"","orcid":"","institution":"University of Glasgow","correspondingAuthor":false,"prefix":"","firstName":"Bernd","middleName":"","lastName":"Porr","suffix":""},{"id":535112719,"identity":"9c77d2df-38ef-46f1-8b03-7214eefcdb60","order_by":3,"name":"Donald C McMillan","email":"","orcid":"","institution":"University of Glasgow","correspondingAuthor":false,"prefix":"","firstName":"Donald","middleName":"C","lastName":"McMillan","suffix":""},{"id":535112720,"identity":"84e49be1-e3b0-4abf-8a55-241f7c24a854","order_by":4,"name":"Ross D Dolan","email":"","orcid":"","institution":"University of Glasgow","correspondingAuthor":false,"prefix":"","firstName":"Ross","middleName":"D","lastName":"Dolan","suffix":""}],"badges":[],"createdAt":"2025-09-10 18:38:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7585523/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7585523/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":94622563,"identity":"06fbfaf4-821b-4a7f-b7bc-8e71f84e2667","added_by":"auto","created_at":"2025-10-29 04:18:22","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":577910,"visible":true,"origin":"","legend":"","description":"","filename":"250916FinalManuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/7c12d5249792111fe4258031.docx"},{"id":94622706,"identity":"1207c5b2-5000-4dde-95ff-642dd4c6146a","added_by":"auto","created_at":"2025-10-29 04:18:29","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7203,"visible":true,"origin":"","legend":"","description":"","filename":"ba9ba91786394912bcafbf87ddf60481.json","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/b9cbe242e3d3838f61fde741.json"},{"id":94622202,"identity":"cc895550-e36f-4076-af35-a2a0354f8969","added_by":"auto","created_at":"2025-10-29 04:18:11","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":80563,"visible":true,"origin":"","legend":"","description":"","filename":"ba9ba91786394912bcafbf87ddf604811enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/700b1b0c6955b06bff329caa.xml"},{"id":94622903,"identity":"631f0469-daf4-4c3f-91b4-5ea730846da0","added_by":"auto","created_at":"2025-10-29 04:18:38","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/f2a6471e27ed23d76283b114.jpeg"},{"id":94622180,"identity":"5b215780-246e-4c0f-94e3-b0ee9c135756","added_by":"auto","created_at":"2025-10-29 04:18:08","extension":"jpeg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/e23018e25073490bcea9dcd7.jpeg"},{"id":94622423,"identity":"bcb7ae03-8aad-4f64-ad3a-e994f68a2c21","added_by":"auto","created_at":"2025-10-29 04:18:18","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/dc03f033cbe1bfc511cb1bd8.jpeg"},{"id":94622618,"identity":"ccf98d5b-be48-43c3-8803-7ba1c977a286","added_by":"auto","created_at":"2025-10-29 04:18:24","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":71629,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/3d9f4affed676738d051164d.png"},{"id":94622494,"identity":"768723a2-c2b4-4a8a-882c-c48380999bd7","added_by":"auto","created_at":"2025-10-29 04:18:21","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":81311,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/47399b3ea63eded1e7fc8a62.png"},{"id":94622717,"identity":"434a90f3-c6b5-4f87-b22f-1b99046c175e","added_by":"auto","created_at":"2025-10-29 04:18:29","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":126308,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/8729f162db8b501b1e671993.png"},{"id":94622210,"identity":"6114e71d-40c9-4723-ba21-57a7291eada3","added_by":"auto","created_at":"2025-10-29 04:18:11","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":116211,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/2342f21f42fb6c27a3d14243.png"},{"id":94622837,"identity":"91bde6bd-d987-43a8-bb0c-e3bf28729f3f","added_by":"auto","created_at":"2025-10-29 04:18:34","extension":"jpeg","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":13267,"visible":true,"origin":"","legend":"","description":"","filename":"groupimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/934c4a9733f035b1b0bba1dc.jpeg"},{"id":94622361,"identity":"110c57d2-5f24-4b15-80a3-449139d334aa","added_by":"auto","created_at":"2025-10-29 04:18:17","extension":"jpeg","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":12214,"visible":true,"origin":"","legend":"","description":"","filename":"groupimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/82372888c21d1f5587819328.jpeg"},{"id":94622247,"identity":"a8c0894b-97fd-4f5d-a756-6d310f8b47d2","added_by":"auto","created_at":"2025-10-29 04:18:13","extension":"jpeg","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":26852,"visible":true,"origin":"","legend":"","description":"","filename":"groupimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/9cd8848b3cef69ba69fa5fc6.jpeg"},{"id":94622565,"identity":"62b12a59-de6a-47b4-ad28-2348aa66b74e","added_by":"auto","created_at":"2025-10-29 04:18:22","extension":"jpeg","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":26026,"visible":true,"origin":"","legend":"","description":"","filename":"groupimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/f4ff26c6a9a093fa2ec6a268.jpeg"},{"id":94622884,"identity":"af18f4c4-1c2f-426f-9edd-a3dad09ae06e","added_by":"auto","created_at":"2025-10-29 04:18:37","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/59249507c47798e458e50af9.png"},{"id":94622623,"identity":"a3e672e9-9de0-4bad-8e1e-6f311f755386","added_by":"auto","created_at":"2025-10-29 04:18:24","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/a241a614215587dd4f5a46d7.png"},{"id":94622380,"identity":"255f87aa-40cb-47c6-8407-c44d0c3b4981","added_by":"auto","created_at":"2025-10-29 04:18:18","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/ba1396ca2e6dfc5061a8e675.png"},{"id":94622536,"identity":"03bbad43-e738-4faf-b142-9ab52b76015e","added_by":"auto","created_at":"2025-10-29 04:18:21","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":12269,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/4f253ffd49972d8cfdcc2763.png"},{"id":94622540,"identity":"b64ebb29-7a47-4c3c-808b-fdb7a7d944f4","added_by":"auto","created_at":"2025-10-29 04:18:22","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15513,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/984301ead8a7290a4493327c.png"},{"id":94622466,"identity":"437e82db-8c06-4a6f-a26a-7c0727e342bc","added_by":"auto","created_at":"2025-10-29 04:18:20","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":23089,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/c62d53d1c539945c928c6cac.png"},{"id":94622711,"identity":"e80c5e58-afee-43ab-a58a-3938a2931391","added_by":"auto","created_at":"2025-10-29 04:18:29","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":22368,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/faa1bad5b431f9e90338087f.png"},{"id":94622353,"identity":"743e00d4-5811-495b-ae9c-18bbc05f1996","added_by":"auto","created_at":"2025-10-29 04:18:17","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3283,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinegroupimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/ef8ef100dea7b3707ad34ccd.png"},{"id":94622747,"identity":"1ae07222-f77a-4fa3-bd1c-a33c67e7e9ce","added_by":"auto","created_at":"2025-10-29 04:18:31","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2753,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinegroupimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/756b960ac28975ff6e13795b.png"},{"id":94622821,"identity":"5a688c69-d119-4f26-9697-10d1e791f15c","added_by":"auto","created_at":"2025-10-29 04:18:33","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5361,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinegroupimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/8eb2095079d201e1193d3c49.png"},{"id":94622360,"identity":"da3eb857-c7f8-4ecb-add8-323ab1d37040","added_by":"auto","created_at":"2025-10-29 04:18:17","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5200,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinegroupimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/deadc42c6853d3ce6bf5a961.png"},{"id":94622659,"identity":"01d45da6-a554-46bd-9161-08eb125f2f4c","added_by":"auto","created_at":"2025-10-29 04:18:26","extension":"xml","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":78632,"visible":true,"origin":"","legend":"","description":"","filename":"ba9ba91786394912bcafbf87ddf604811structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/fc9c252d303eeee30145f796.xml"},{"id":94622430,"identity":"87ca63d7-bd44-4602-bcf7-2ae5de629ec4","added_by":"auto","created_at":"2025-10-29 04:18:19","extension":"html","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":87717,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/13971d7438377d29f7860945.html"},{"id":94622530,"identity":"e51d2df4-38e9-4699-8b53-0ba265f78153","added_by":"auto","created_at":"2025-10-29 04:18:21","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":27121,"visible":true,"origin":"","legend":"\u003cp\u003eMean percentage error (PE) by model (1a) and by body composition metric (1b). GoogLeNet and AlexNet achieved the lowest overall PE compared with UNet and ResNet34.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/ac1c253debd195e888669c30.png"},{"id":94622851,"identity":"8a525897-40aa-4924-8c39-7c13d74cce81","added_by":"auto","created_at":"2025-10-29 04:18:35","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":37487,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance of GoogLeNet in predicting skeletal muscle area (SMA) across epochs (2a), and performance of AlexNet in predicting skeletal muscle density (SMD) across epochs (2b).\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/f038a3a824bb5ad844c4a7db.png"},{"id":94622532,"identity":"1b4697d8-209a-4365-a56b-9f95e0d384f3","added_by":"auto","created_at":"2025-10-29 04:18:21","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":86253,"visible":true,"origin":"","legend":"\u003cp\u003eScatter plots showing predicted versus true values for skeletal muscle index (SMI) (3a) and skeletal muscle density (SMD) (3b).\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/04dc2dc57700be64ea8aba9e.png"},{"id":94622891,"identity":"e16c0768-73e8-4a23-b409-37ac9b8ff7d7","added_by":"auto","created_at":"2025-10-29 04:18:38","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":113501,"visible":true,"origin":"","legend":"\u003cp\u003eLinear regression analysis between predicted and true values for (a) skeletal muscle index (SMI) and (b) skeletal muscle density (SMD).\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/369c7f0f39775ba74aab7b3a.png"},{"id":94623801,"identity":"fc05452e-9076-4242-8668-2b0e092ab902","added_by":"auto","created_at":"2025-10-29 04:19:48","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":836748,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7585523/v1/535aeb13-7728-4ded-a540-06baffed6556.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eDeep learning-based computed tomography (CT) derived body composition classifier for colorectal cancer patients\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003eComputed Tomography (CT) scans have emerged as invaluable tools in cancer management, providing detailed insights into patient anatomy that are crucial for prognosis and treatment planning. However, the full prognostic potential of CT-derived body composition analysis remains underutilized in clinical practice. This underutilization is largely due to the time-consuming nature of the analysis and the specialized expertise required to accurately interpret the data (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). Furthermore, the lack of access to advanced analytical tools, the high costs associated with these technologies, and insufficient training among healthcare professionals further limit the widespread adoption of these critical assessments (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). Historically, Body Mass Index (BMI) has been the standard metric used for assessing the nutritional and physical status of cancer patients. However, with the increasing prevalence of obesity, BMI has been shown to be an inadequate measure for predicting outcomes in oncology. Unlike BMI, CT-derived metrics provide a more nuanced and clinically relevant assessment of a patient\u0026rsquo;s body composition (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). These metrics include skeletal muscle area (SMA), skeletal muscle density (SMD), subcutaneous fat area (SFA), and visceral fat area (VFA), all of which offer critical insights into a patient\u0026rsquo;s health status.\u003c/p\u003e\u003cp\u003eThe importance of these metrics extends beyond simple body composition. For example, sarcopenia, which is characterized by the degenerative loss of muscle mass and function, is a common condition in cancer patients that is associated with poor clinical outcomes, including reduced survival rates (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e). Similarly, cachexia, which involves a disproportionate loss of skeletal muscle relative to fat, significantly impacts a patient\u0026rsquo;s quality of life and overall prognosis (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e). CT-derived metrics such as SMD can also indicate myosteatosis, a condition where muscle tissue is infiltrated by fat, further complicating a patient's clinical outlook.\u003c/p\u003e\u003cp\u003eOne of the most significant CT-derived metrics is the Skeletal Muscle Index (SMI), which normalizes the SMA to a patient's height squared, providing a standardized measure of muscle mass that is strongly associated with survival in cancer patients. When combined with SFA, VFA, and SMD, these metrics can be used to calculate a CT-derived sarcopenia score (CT-SS), which serves as a powerful prognostic tool. The CT-SS is categorized into three levels, with higher scores indicating poorer prognosis. Despite their proven utility, the application of these measures in routine clinical practice remains limited due to the labour-intensive nature of manual analysis and the complexity of integrating these metrics into clinical workflows (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThe advent of deep learning technologies offers a promising solution to these challenges. Deep learning, particularly convolutional neural networks (CNNs), has demonstrated remarkable success in automating complex image analysis tasks, making it feasible to incorporate detailed body composition assessments into routine clinical practice. By leveraging large datasets, these networks can learn to accurately predict key body composition metrics from CT scans, potentially transforming the way clinicians assess and manage cancer patients (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e). This study aims to evaluate the effectiveness of several deep learning architectures, including AlexNet (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e), UNet (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e), GoogLeNet (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e), and ResNet34 (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e) in predicting body composition metrics from CT scans of colorectal cancer patients. The ultimate goal is to develop a robust, clinically applicable tool that can enhance prognostic accuracy and improve patient outcomes in oncology.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cb\u003eData Preparation\u003c/b\u003e: This study utilized two primary datasets: one consisting of CT scans from the third lumbar vertebra (L3) of patients diagnosed with colorectal cancer at the Glasgow Royal Infirmary between 2008 and 2018, and another comprising anonymized patient information in a tabulated format. Initially, the CT scans were in DICOM format, which was incompatible with TensorFlow, the deep learning framework used in this study. To address this, the images were batch-converted to JPEG format using ImageJ (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). To optimize computational efficiency, the images were resized from 512\u0026times;512 pixels to 256\u0026times;256 pixels. A total of 574 CT scans were retained after processing, followed by a visual inspection for anisometry errors. Ten scans were randomly selected and set aside as a prediction set for post-training model evaluation.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eModel Selection and Evaluation\u003c/strong\u003e\u003cp\u003eFour well-established deep learning architectures were selected for this study. AlexNet (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e) was chosen as a benchmark due to its simplicity and efficiency, achieved through the use of convolutional layers, ReLU activations, and Max-Pooling layers. These features make AlexNet particularly well-suited for handling large-scale image data with relatively low computational costs. UNet (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e) was included for its strength in biomedical image segmentation, especially its ability to accurately extract detailed features from images, which is critical for segmenting body composition metrics like SFA and VFA from CT scans. GoogLeNet (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e) was selected for its innovative inception modules, which enable the model to capture multi-scale features within a single layer. This capability makes it particularly effective in distinguishing between different tissue types in complex CT images, while its use of global average pooling helps prevent overfitting. ResNet34 (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e) was chosen for its depth and the use of residual connections, which help mitigate the vanishing gradient problem in deep networks, making it suitable for capturing intricate patterns in CT scan data, particularly for SMD and SMA predictions.\u003c/p\u003e\u003c/p\u003e\u003cp\u003eThe models were evaluated based on their performance in predicting SFA, VFA, SMA, and SMD, using Mean Squared Error (MSE) and Percentage Error (PE) as metrics. Given that these output variables are continuous, regression models were constructed for each architecture. The training process commenced with the setup of TensorFlow and Keras, the core libraries used for developing and fine-tuning the deep learning models. The CT scan data were carefully pre-processed and organized into a structured dataset, which included converting DICOM images to JPEG format and resizing them to 256\u0026times;256 pixels to optimize for computational efficiency. Hyperparameters, including batch size of 16 (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e), learning rate (0.0001), and epochs (100), were uniformly initialized across all models. A seed value of 42 was used to ensure reproducibility.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eData Augmentation and Training\u003c/strong\u003e\u003cp\u003eTo improve the models' ability to generalize, data augmentation techniques such as random flips, rotations, and zooms were applied. These augmentations aimed to reduce overfitting by introducing variability into the training data. The models were built using the Keras functional API, which provides flexibility in model architecture design. An 80/20 data split was used for training and validation, with the MSE recorded at each epoch as the loss function.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eModel Optimization\u003c/strong\u003e\u003cp\u003eFollowing the initial training, GoogLeNet demonstrated the highest accuracy in predicting SMA, making it the focus of further optimization using Hyperband tuning (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e). Hyperband was selected for its efficiency in optimizing hyperparameters under limited computational resources. The key hyperparameters optimized for GoogLeNet included filter numbers, learning rate, and seed. Similarly, AlexNet, which performed well in predicting SMD, underwent optimization focusing on dropout rate, dense layer nodes, and learning rate.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eDevelopment of the CT-SS Classifier\u003c/strong\u003e\u003cp\u003ePost-optimization, the most accurate models\u0026mdash;GoogLeNet for SMA and AlexNet for SMD\u0026mdash;were used to develop a CT-SS classifier. The CT-SS was calculated using the SMI, derived by normalizing SMA by the square of the patient\u0026rsquo;s height. SMI and SMD values were classified as low or not low based on established thresholds dependent on BMI, sex, and height. The final CT-SS score, ranging from 0 to 2, was computed, with scores of 2 indicating low SMI and SMD, 1 indicating low SMI or SMD, and 0 indicating normal or high values in both metrics.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eWeb Application Development\u003c/strong\u003e\u003cp\u003eTo enable the practical application of the developed models in clinical settings, a web-based application was created using Streamlit. This app allows clinicians to input patient data, including sex, height, BMI, and upload CT scans for analysis. The app predicts SMA and SMD using the trained models, calculates SMI, and provides a CT-SS score. The user-friendly interface ensures that the tool can be seamlessly integrated into clinical practice, facilitating quick and accurate assessments.\u003c/p\u003e\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cb\u003eModel Performance\u003c/b\u003e: The performance of the four deep learning models was evaluated based on their ability to predict SFA, VFA, SMA, SMD from CT scans. The evaluation metrics used included MSE and PE. AlexNet excelled in predicting SMD, achieving a PE of 8.12% and an MSE of 40. For SMA, AlexNet recorded a PE of 6.15% and an MSE of 468. UNet showed competitive performance, particularly in predicting SFA and VFA, with a PE of 8.76% (MSE: 3384) for SFA and a PE of 9.52% (MSE: 3998) for VFA. GoogLeNet demonstrated the strongest overall performance, achieving the lowest PE across multiple metrics. Specifically, GoogLeNet predicted SMA with a PE of 4.96% and an MSE of 450. For SMD, it achieved a PE of 8.04% and an MSE of 34. ResNet34 recorded a PE of 10.30% for SFA (MSE: 3493) and a PE of 10.56% for SMD (MSE: 40). It particularly struggled with predicting VFA, where it had a PE of 12.02% and an MSE of 5935. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e summarizes these results. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e1\u003c/span\u003ea illustrates the PE for each model across all body composition metrics, highlighting that GoogLeNet achieved the lowest overall PE at 8.04%, followed closely by AlexNet at 8.52%. In contrast, UNet and ResNet34 had higher PEs at 10.30% and 11.56%, respectively, while Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e1\u003c/span\u003eb illustrates the mean PE for each body composition metric. The architectures encountered the greatest difficulty with SFA, with a mean PE of 10.9%. The best performance was with SMA, making it the most accurate predicted overall with a mean PE of 7.29%.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison of the Different Models\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"13\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c12\" colnum=\"12\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c13\" colnum=\"13\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e\u003cp\u003eAlexNet\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"3\" nameend=\"c7\" namest=\"c5\"\u003e\u003cp\u003eGoogLeNet\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"3\" nameend=\"c10\" namest=\"c8\"\u003e\u003cp\u003eUNet\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"3\" nameend=\"c13\" namest=\"c11\"\u003e\u003cp\u003eResNet34\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMSE\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAE\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003ePE%\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eMSE\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eAE\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003ePE%\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eMSE\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eAE\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003ePE%\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c11\"\u003e\u003cp\u003eMSE\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c12\"\u003e\u003cp\u003eAE\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c13\"\u003e\u003cp\u003ePE%\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSFA (cm\u0026sup2;)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e3412\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e58.41\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e8.16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e6869\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e82.88\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e11.58\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e5935\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e77.04\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e10.76\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e9896\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e99.48\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c13\"\u003e\u003cp\u003e13.90\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eVFA (cm\u0026sup2;)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e3998\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e63.23\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e9.52\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e1498\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e38.70\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e5.83\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e3493\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e59.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e8.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e3384\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e58.17\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c13\"\u003e\u003cp\u003e8.76\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSMA (cm\u0026sup2;)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e468\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e21.63\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e6.27\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e450\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e21.21\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e6.15\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e882\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e29.70\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e8.60\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e868\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e29.46\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c13\"\u003e\u003cp\u003e8.53\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSMD (HU)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e34\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e5.83\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e9.28\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e40\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e6.32\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e10.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e57\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e7.55\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e12.02\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c11\"\u003e\u003cp\u003e44\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e6.63\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c13\"\u003e\u003cp\u003e10.56\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eTime Efficiency\u003c/strong\u003e\u003cp\u003eThe computational experiments were performed using an Intel(R) Xeon(R) CPU E5630 @ 2.53GHz and a GeForce GTX 1070 GPU. The combination of these hardware components allowed for efficient processing of the deep learning models. The time required per epoch for each model was recorded to assess computational efficiency. AlexNet completed each epoch in 1 second, making it the fastest model, GoogLeNet completed each epoch in 3 second, UNet completed each epoch in 12 second, whereas ResNet34 required 60 seconds per epoch, making it the slowest.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eHyperparameter Optimization\u003c/strong\u003e\u003cp\u003eHyperband hyperparameter tuning was applied to optimize the GoogLeNet model for SMA and the AlexNet model for SMD. After optimization, GoogLeNet's SMA prediction error was reduced from 6.15% to 5.65%, and AlexNet's SMD prediction error decreased from 9.28% to 8.96%.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eModel Customization and Final Performance\u003c/strong\u003e\u003cp\u003eFurther customization of the models was performed to enhance their performance. Figure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e2\u003c/span\u003ea illustrates the change in PE when increasing the number of epochs for GoogLeNet with SMA. Although 1000 epochs exhibited the lowest PE, it demonstrated a high degree of unreliability and experienced interruptions during training due to out of memory (OOM) errors. A local minimum is found at 300 epochs that was consistent across multiple reruns and had no reliability issues. Given the selection of 300 epochs for GoogLeNet. Figure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e2\u003c/span\u003eb shows the change in PE for AlexNet training with SMD at different epochs. The lowest PE value, 8.12%, achieved was at 600 epochs. there were no OOM errors encountered during training. Given the selection of 600 epochs for AlexNet, these values were determined through iterative testing and careful analysis of the models' learning curves. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e summarises the overall impact of these changes on the performance of the GoogLeNet and AlexNet models.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eImpact of Changes implemented\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eChange Made\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e\u003cp\u003eGoogLeNet SMA Model\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c5\" namest=\"c4\"\u003e\u003cp\u003eAlexNet SMD Model\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePE% Achieved\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eImprovement %\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003ePE% Achieved\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eImprovement %\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eOriginal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e6.15\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e9.28\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHyperBand Optimizing\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e5.65\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-0.50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e8.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-0.32\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAlternative End\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e5.65\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e8.28\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-0.68\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNumber of Epochs\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e4.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-0.69\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e8.12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-0.16\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBatch Size\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e4.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e8.61\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTotal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1.19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1.16\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eCT-SS Classifier Results\u003c/strong\u003e\u003cp\u003eThe CT-SS classifier, developed using the optimized GoogLeNet and AlexNet models, was tested on a separate set of CT scans. The mean absolute error for SMI predictions was 2.299, with a relative mean error of 5.02% Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e3\u003c/span\u003ea. The regression equation indicates that the associated p-value\u0026thinsp;=\u0026thinsp;0.0005 for SMI predictions and true SMI Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e4\u003c/span\u003ea. For SMD predictions the mean absolute error was 1.852, resulting in a relative mean error of 9.85% Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e3\u003c/span\u003eb. The regression equation indicates that the associated p-value\u0026thinsp;=\u0026thinsp;0.0004 for SMD predictions and true SMD Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e4\u003c/span\u003eb. The classifier achieved an overall accuracy of 80% for the CT-SS classifications.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eWeb Application Deployment\u003c/strong\u003e\u003cp\u003eA Model application was developed to integrate the trained models into a user-friendly interface for clinical use. The application allows users to input patient data, including sex, height, and BMI, and upload CT scans to receive predictions for SMA, SMD, and the CT-SS score.\u003c/p\u003e\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe deep learning models employed in this study were selected based on their balance between accuracy and computational efficiency, which are critical factors for practical clinical application. Among the architectures tested, GoogLeNet emerged as the most suitable for predicting SMA, while AlexNet was the top performer for SMD. These selections were driven by their performance metrics and computational demands during training. GoogLeNet's superior performance can be attributed to its innovative use of parallel layers with multiple filter sizes, which allows the model to process input data at various scales within a single layer. This design reduces the computational burden by incorporating global average pooling layers, which minimize the number of trainable parameters and help prevent overfitting. Additionally, the reliance on small, computationally inexpensive convolutions further enhances GoogLeNet's efficiency, making it comparable to AlexNet despite its deeper architecture.\u003c/p\u003e\u003cp\u003eAlexNet, although simpler and shallower than the other models, demonstrated the highest accuracy in predicting SMD. This is likely due to its compact architecture, which reduces the overall number of trainable parameters and accelerates the optimization process. The use of Max-Pooling and dropout layers further contributes to its computational efficiency by limiting the number of nodes and computations required per epoch. Interestingly, the relatively shallow depth of AlexNet may have allowed it to avoid overfitting on non-relevant features, thus maintaining high accuracy in SMD prediction.\u003c/p\u003e\u003cp\u003eConversely, ResNet34, the deepest model tested, underperformed both in terms of accuracy and computational efficiency. The absence of layers designed to reduce computational costs, such as Max-Pooling or Global Average Pooling (GAP), combined with its substantial depth, likely contributed to its lower efficiency and accuracy. This underperformance suggests that ResNet34 may have encountered degradation issues, where the added depth did not translate into improved accuracy, potentially due to noise or irrelevant patterns in the data being learned by the model.\u003c/p\u003e\u003cp\u003eThe prediction accuracy varied across the body composition metrics, with SMA and SMD being predicted more accurately than SFA and VFA. The PE for SMA was notably lower, especially with the GoogLeNet model, which achieved a PE of 6.15%. This performance underscores the model's robustness in capturing muscle-related features from CT scans. In contrast, SFA predictions were the least accurate, with a mean PE of 10.9%, proving particularly challenging for both GoogLeNet and ResNet34. The difficulty in distinguishing between SFA and VFA, given their similar appearances on CT scans, may have contributed to this higher error rate. Additionally, the edge detection required to separate SFA from the background might have been a limiting factor for these models. Despite these challenges, VFA was predicted with slightly better accuracy, indicating that the models could more reliably detect fat distributions closer to the visceral organs. SMD predictions, while more accurate than SFA, still presented challenges, particularly due to the nature of the metric, which relies on the brightness of the muscle in Hounsfield units. The percentage error for SMD averaged 10.44%, with AlexNet achieving the lowest error at 9.28%. This suggests that while the models are proficient at detecting muscle mass, further refinement is necessary to accurately capture muscle density.\u003c/p\u003e\u003cp\u003eThe Hyperband optimization method proved effective in enhancing model performance, particularly for the GoogLeNet SMA model, where the percentage error was reduced by 0.459%. This improvement was achieved by fine-tuning the initial and final filter numbers, which increased the number of trainable parameters, leading to better accuracy without compromising computational efficiency. The learning rate, a critical parameter for balancing convergence speed and precision, was also optimized, contributing to the model's improved performance. For the AlexNet SMD model, the Hyperband tuning resulted in a smaller but significant improvement of 0.32%. The optimization focused on the dense and dropout layers, with the most substantial gain in accuracy (0.68%) achieved by replacing the fully connected, dense layers with those from GoogLeNet. This substitution not only enhanced accuracy but also reduced computational costs, demonstrating the effectiveness of combining elements from different architectures to leverage their strengths. The most considerable improvement for the GoogLeNet SMA model came from increasing the number of training epochs, which reduced the percentage error by 0.69%. However, the relationship between epoch count and error reduction was non-linear, with diminishing returns observed beyond certain thresholds, likely due to overfitting or noise in the dataset being misinterpreted as relevant features. Similar trends were observed for AlexNet, where the percentage error decreased up to 600 epochs but increased slightly thereafter, likely due to overfitting.\u003c/p\u003e\u003cp\u003eThe CT-SS classifier, developed using the optimized GoogLeNet and AlexNet models, demonstrated an overall accuracy of 80% in classifying SMI and SMD as \"low\" or \"not low.\" The classifier's performance, as indicated by the F1 score of 0.75, was reasonably balanced, though there is room for improvement, particularly in reducing false positives. The errors observed in classification were concentrated in specific patient subgroups, notably female patients for SMI and male patients with a BMI less than 25 for SMD. These findings suggest that further research is needed to explore sex- and BMI-related variations in model accuracy, potentially requiring tailored approaches for different demographic groups.\u003c/p\u003e\u003cp\u003eThe developed web application served as a proof of concept for integrating deep learning models into a user-friendly clinical tool. The app's performance was consistent, with predictions generated in under a second per scan, and the CT-SS predictions matched those calculated manually from the predicted values. However, the app's reliance on correct input data format and the lack of safeguards against erroneous inputs are notable limitations. Despite these issues, the app's intuitive interface and robust performance in handling valid data make it a promising tool for clinical use, provided further refinements are made to address its current limitations.\u003c/p\u003e\u003cp\u003eIn comparing our results with existing work in the field, we found that several studies have applied deep learning models to body composition analysis, particularly using abdominal CT scans. For instance, Hsu et al. (2021) developed a machine learning model to predict body composition metrics such as visceral fat area (VFA), subcutaneous fat area (SFA), and skeletal muscle area (SMA) from CT scans of patients with pancreatic cancer. However, their approach differs from ours in several key aspects: their study utilized a U-Net architecture to perform image segmentation (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e), whereas our work adopts a black-box approach, providing a numerical output for body composition metrics, including skeletal muscle density (SMD). Additionally, our approach achieved a superior performance, particularly for SMA prediction, where GoogLeNet attained a percentage error (PE) of 6.15%.\u003c/p\u003e\u003cp\u003eSimilarly, Dabiri et al. (2019) applied machine learning techniques to segment muscles and fat tissues from CT scans at the lumbar (L3) and thoracic (T4) levels in patients with colorectal cancer. Their approach also relied on image segmentation, focusing on muscle segmentation at L3, like our study (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e). However, our work distinguishes itself by focusing on both segmentation and the prediction of numerical metrics, such as SMD and fat area, with AlexNet showing the highest accuracy for SMD prediction, achieving a PE of 9.28%. Although Dabiri et al. did not provide detailed performance metrics comparable to ours, it is clear that our black-box numerical prediction approach offers a novel contribution, particularly for SMD prediction, which remains an underexplored area.\u003c/p\u003e\u003cp\u003ewhile previous work has primarily focused on image segmentation for body composition analysis, our study presents a novel approach by providing numerical outputs for metrics such as SMA, SFA, VFA, and SMD. Our results indicate that the deep learning models tested, particularly GoogLeNet and AlexNet, offer competitive and in some cases superior accuracy to existing segmentation-based approaches, particularly in the prediction of SMA and SMD. This highlights the practical applicability of our approach in clinical settings, where rapid and accurate predictions of body composition metrics are essential for assessing patient health, especially in individuals with conditions like colorectal cancer.\u003c/p\u003e\u003cp\u003eIn conclusion, the results of this study demonstrate the feasibility of using deep learning models, specifically GoogLeNet and AlexNet, for accurate body composition analysis in clinical settings. The optimization processes applied have significantly enhanced the models' performance, making them more suitable for practical application. The study underscores the potential of these deep learning models to automate the analysis of CT-derived body composition metrics, which are crucial for assessing patient health, particularly in individuals with colorectal cancer. By accurately predicting these metrics, the models can assist clinicians in making informed decisions regarding patient prognosis, treatment planning, and monitoring the progression of conditions such as sarcopenia and cachexia. The models' ability to efficiently handle large-scale image data positions them as viable candidates for integration into clinical workflows.\u003c/p\u003e\u003cp\u003eHowever, the study also highlights several areas for future research. There remains room for further optimization, particularly in improving the prediction accuracy of SFA and VFA, where the models exhibited higher PEs. Future work could explore the use of more advanced data augmentation techniques or the incorporation of transfer learning to further enhance model performance. Additionally, expanding the dataset to include a more diverse range of patient demographics and health conditions could improve the models' generalization capabilities. Moreover, real-world clinical validation is essential to ensure their robustness and reliability in everyday practice. The development of a user-friendly web application, as outlined in the study, represents a significant step toward this goal by providing clinicians with a practical tool for quick and accurate body composition analysis.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003eEthics Statement\u003c/h2\u003e\u003cp\u003eThis study was reviewed and approved by the West of Scotland Research Ethics Committee, Glasgow. All research was conducted in accordance with the Declaration of Helsinki. Written informed consent for participation was waived in accordance with national legislation and the requirements of the West of Scotland Research Ethics Committee because the study involved retrospective analysis of anonymized data collected during routine clinical care.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003ch2\u003e\u003cb\u003eConsent for publication\u003c/b\u003e:\u003c/h2\u003e\u003cp\u003eNot applicable.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003ch2\u003eCompeting Interests:\u003c/h2\u003e\u003cp\u003eThe author(s) declare no conflict of interest.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFunding:\u003c/h2\u003e\u003cp\u003eThe author(s) received no specific funding for this work.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eEH designed the work, acquired data, interpreting the results, revised the manuscript, Approved the final version. CP interpreting the results, drafted or revised the manuscript, Approved the final version. BP designed the work, revised the manuscript, Approved the final version. DCM designed the work, acquired data, interpreting the results, revised the manuscript, Approved the final version. RDD designed the work, acquired data, interpreting the results, revised the manuscript, Approved the final version.\u003c/p\u003e\u003ch2\u003eAcknowledgements:\u003c/h2\u003e\u003cp\u003eThe authors have nothing to report.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets generated and/or analysed during the current study are available from the corresponding author on reasonable request. The source code used for the analysis is publicly available at Zenodo ( [https://zenodo.org/records/10439107](https:/zenodo.org/records/10439107) ) (18)\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eMcGovern J, Delaney J, Forshaw MJ, McCabe G, Crumley AB, McIntosh D, et al. The relationship between computed tomography-derived sarcopenia, cardiopulmonary exercise testing performance, systemic inflammation, and survival in good performance status patients with oesophago-gastric cancer undergoing neoadjuvant treatment. JCSM Clin Rep. 2023;8(1):3\u0026ndash;11.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMcGovern J, Golder AM, Dolan RD, Roxburgh CSD, Horgan PG, McMillan DC. The combination of computed tomography-derived muscle mass and muscle density and relationship with clinicopathological characteristics and survival in patients undergoing potentially curative surgery for colorectal cancer. JCSM Clin Rep. 2022;7(3):65\u0026ndash;76.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eArends J, Baracos V, Bertz H, Bozzetti F, Calder PC, Deutz NEP, et al. ESPEN expert group recommendations for action against cancer-related malnutrition. Clin Nutr. 2017;36(5):1187\u0026ndash;96.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTolonen A, Pakarinen T, Sassi A, Kytta J, Cancino W, Rinta-Kiikka I, et al. Methodology, clinical applications, and future directions of body composition analysis using computed tomography (CT) images: A review. Eur J Radiol. 2021;145:109943.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVergara-Fernandez O, Trejo-Avila M, Salgado-Nesme N. Sarcopenia in patients with colorectal cancer: A comprehensive review. World J Clin Cases. 2020;8(7):1188\u0026ndash;202.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFearon K, Strasser F, Anker SD, Bosaeus I, Bruera E, Fainsinger RL, et al. Definition and classification of cancer cachexia: an international consensus. Lancet Oncol. 2011;12(5):489\u0026ndash;95.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSomasundaram E, Castiglione JA, Brady SL, Trout AT. Defining Normal Ranges of Skeletal Muscle Area and Skeletal Muscle Index in Children on CT Using an Automated Deep Learning Pipeline: Implications for Sarcopenia Diagnosis. AJR Am J Roentgenol. 2022;219(2):326\u0026ndash;36.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDolan RD, Almasaudi AS, Dieu LB, Horgan PG, McSorley ST, McMillan DC. The relationship between computed tomography-derived body composition, systemic inflammatory response, and survival in patients undergoing surgery for colorectal cancer. J Cachexia Sarcopenia Muscle. 2019;10(1):111\u0026ndash;22.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKrizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Commun ACM. 2017;60(6):84\u0026ndash;90.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRonneberger O, Fischer P, Brox T, editors. U-Net: Convolutional Networks for Biomedical Image Segmentation. Cham: Springer International Publishing; 2015.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSzegedy C, Liu W, Jia Y, Sermanet P, Reed SE, Anguelov D et al. Going deeper with convolutions. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2014:1\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHe K, Zhang X, Ren S, Sun J. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2015:770-8.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSchneider CA, Rasband WS, Eliceiri KW. NIH Image to ImageJ: 25 years of image analysis. Nat Methods. 2012;9(7):671\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKingma DP, Ba J, Adam. A Method for Stochastic Optimization. CoRR. 2014;abs/1412.6980.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi L, Jamieson KG, DeSalvo G, Rostamizadeh A, Talwalkar A. Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization. J Mach Learn Res. 2016;18:1851.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHsu TH, Schawkat K, Berkowitz SJ, Wei JL, Makoyeva A, Legare K, et al. Artificial intelligence to assess body composition on routine abdominal CT scans and predict mortality in pancreatic cancer- A recipe for your local application. Eur J Radiol. 2021;142:109834.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDabiri S, Popuri K, Cespedes Feliciano EM, Caan BJ, Baracos VE, Beg MF. Muscle segmentation in axial computed tomography (CT) images at the lumbar (L3) and thoracic (T4) levels for body composition analysis. Comput Med Imaging Graph. 2019;75:47\u0026ndash;55.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHarling E. 2451792h_MastersProject [Model]. Zenodo: University of Glasgow; 2023.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Deep learning, Computed tomography (CT), Body composition analysis, Colorectal cancer, Medical imaging","lastPublishedDoi":"10.21203/rs.3.rs-7585523/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7585523/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eAccurate 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.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eFour 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.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eGoogLeNet 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.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e\u003cp\u003eOptimized deep learning models, particularly GoogLeNet and AlexNet, can automate CT-derived body composition analysis with high accuracy. 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.\u003c/p\u003e","manuscriptTitle":"Deep learning-based computed tomography (CT) derived body composition classifier for colorectal cancer patients","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-29 03:58:33","doi":"10.21203/rs.3.rs-7585523/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-02-12T09:30:08+00:00","index":"","fulltext":""},{"type":"reviewerAgreed","content":"6354464008681476862263492376308928967","date":"2026-02-12T06:16:43+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-12T04:50:41+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"43658638506520902800251644061084327406","date":"2026-02-01T17:05:00+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-12T21:38:57+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"73654851669661628589676612442712744662","date":"2026-01-12T21:34:43+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"116737467244026336350047246171501825163","date":"2025-12-10T05:05:42+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"154527559543294491069733516079367982865","date":"2025-11-10T01:45:31+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-10-14T01:08:55+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-10-09T00:58:22+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-09-23T11:25:41+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-09-22T12:19:45+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2025-09-16T11:29:37+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a626af5b-f746-4cfe-b86f-96d40940dfe9","owner":[],"postedDate":"October 29th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-21T14:25:04+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-29 03:58:33","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7585523","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7585523","identity":"rs-7585523","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}