The development of an attention mechanism enhanced deep learning model and its application for body composition assessment with L3 CT images

Abstract

Body composition assessment is very useful for evaluating a patient’s status in the clinic, but recognizing, labeling, and calculating the body compositions would be burdensome. This study aims to develop a web-based service that could automate calculating the areas of skeleton muscle (SM), visceral adipose tissue (VAT), and subcutaneous adipose tissue (SAT) according to L3 computed tomography (CT) images. 1500 L3 CT images were gathered from Xuzhou Central Hospital. Of these, 70% were used as the training dataset, while the remaining 30% were used as the validating dataset. The UNet framework was combined with attention gate (AG), Squeeze and Excitation block (SEblock), and Atrous Spatial Pyramid Pooling (ASSP) modules to construct the segmentation deep learning model. The model’s efficacy was externally validated using two other test datasets with multiple metrics, the consistency test and manual result checking. A graphic user interface was also created and deployed using the Streamlit Python package. The custom deep learning model named L3 Body Composition Segmentation Model (L3BCSM) was constructed. The model’s Median Dice is 0.954(0.930, 0.963)(SATA), 0.849(0.774,0.901)(VATA), and 0.920(0.901, 0.936)(SMA), which is equal to or better than classic models, including UNETR and AHNet. L3BCSM also achieved satisfactory metrics in two external test datasets, consistent with the qualified label. An internet-based application was developed using L3BCSM, which has four functional modules: population analysis, time series analysis, consistency analysis, and manual result checking. The body composition assessment application was well developed, which would benefit the clinical practice and related research.

Introduction

Body composition assessment has been a crucial area of research in medicine and health sciences for a long time. It involves analyzing different tissues and substances in the body, such as muscle, adipose, bone, and organs, to quantify their composition. Body composition assessment is used to monitor the effectiveness of treatment, understand the mechanism of diseases, and create personalized medical plans. Jie Lee et al.¹ confirmed that the loss of human skeletal muscle over time is associated with a negative long-term prognosis for cervical cancer patients. Similarly, Amandine Crombée et al.² found that changes in body composition are linked to the response to immunotherapy in patients with metastatic solid tumors. Our research has shown that changes in skeletal muscle content six months after surgery have a significant impact on both 5-year overall survival (OS) and disease-free survival (DFS) for patients with either sarcopenia or non-sarcopenia before surgery³. This was discovered through a 5-year follow-up of 314 colorectal cancer patients who underwent radical surgery. Traditionally, measuring skeletal muscle and adipose areas has relied on manual labeling with software, which significantly limits related research and clinical practice in body composition assessment. The main reason is that manual labeling is a professional-required, time-consuming, heterogeneous, and expensive process that cannot meet the need for dynamic tracking and follow-up in the clinic⁴.

Automatically identifying and segmenting different tissue types in medical images with deep learning methods has raised great interest among researchers^5,6, especially in computed tomography (CT) images^4,7,8. In addition to using the classic Unet network for traditional segmentation tasks on individual CT images above, an advanced 3D-Unet algorithm has been developed for segmenting muscles across the entire body⁹, and 2D Unet has been used for segmenting Erector Spinae Muscles¹⁰ and sternocleidomastoid and skeletal muscles¹¹. Assessing body composition with deep learning models is fast, stable, and easy to operate. Several studies have reported qualified deep-learning models that could perform well in the automatic assessment of skeletal muscle and adipose tissue area^8,12. Benefiting from the introduction of the attention mechanism, researchers tried to integrate the attention mechanism into the classic Unet network to enhance the efficiency of the segmentation model, making it widely used in this area. Multiple attention mechanisms have been combined into a single model in some cases. For example, the original Unet model now includes Attention Gating Modules to focus on regions of interest without sacrificing spatial resolution and improve feature maps’ quality. A Squeeze and Excitation block (SEblock) is also placed between the last encoder and the first decoder¹³. Similarly, a novel dense residual U-Net model combines attention gate (AG), efficient channel attention (ECA), and Atrous Spatial Pyramid Pooling (ASPP) to improve the performance of automatic segmentation of multiple sclerosis (MS) lesions using 3D Magnetic Resonance Imaging (MRI) sequences¹⁴. Building on this concept, our study aimed to integrate three types of attention mechanisms (AG, SE Block, and ASSP) into a single model for the segmentation of CT images. With the fast development of deep learning algorithms, such as space and multi-scale fusion attention, developing a better performance model for body composition assessment is possible.

Besides this, more and more researchers are inclined to develop web applications to showcase their models^15,16; therefore, users could evaluate the model based not only on the article’s description but also on its actual functioning. While most web apps only include a simple module to display the model’s output, this may not guarantee that users can effectively apply the model in clinical practice. In this article, we introduce and discuss the essential modules required for a model intended for use in clinical practice. This work will facilitate the model’s seamless integration into clinical workflows, aiming to promote its adoption in clinical practice.

Methods

Data collection

Abdominal L3 CT scans of 1500 patients were collected from three hospitals across different regions and periods, from 2018 to 2022. To ensure dataset diversity, we intentionally sought collaboration with institutions that serve diverse patient populations, including healthy patients, kinds of abdominal cancer patients, and sepsis patients. The dataset was divided into training and validation datasets in a 7:3 ratio. Additionally, there were two external test datasets in this study. Of the two external test data, one was collected from the Northern Theater General Hospital; the other dataset was obtained from Eastern Theater General Hospital (Table 1, Fig. 1). The Xuzhou Central Hospital Ethics Committee waived the need for informed consent from all subjects and/or their legal guardians according to the local legislation of Measures for the Ethical Review of Life Science and Medical Research Involving Humans (2023)¹⁷. The Xuzhou Central Hospital Ethics Committee approved the research according to a research guideline for artificial intelligence in medical imaging (CLIAM)¹⁸ (XZYY-LJ-20210715–060). All methods were performed in accordance with the relevant guidelines and regulations. All data was desensitized and encrypted during storage and transmission. Access controls were implemented, and regular security audits were conducted to ensure patients’ privacy.

Table 1.

Characteristics of training and external testing datasets (mean ± SD/n).

Characteristics	Cohort 1 (training and validation)	Cohort 2 (test)	Cohort 3 (test)
Number of scans	1500	160	193
Age (years)	59.63 12.92	57.82 11.54	64.25 12.66
Male, n (%)	1012(67.47%)	97(60.63%)	107(55.44%)
Height (cm)	165.42 8.31	165.56 8.62	149.76 8.54
Weight (kg)	60.51 9.29	58.61 9.37	52.34 6.28
Enhanced/plain	Enhanced/plain	Enhanced/plain	Enhanced/plain
Underlying disease (n)
None (healthy)	200	10	10
Gastric cancer	250	30	20
Colorectal cancer	250	30	20
Live cancer	200	30	20
Ovarian cancer	250	30	20
Lung cancer	200	30	20
Intra-abdominal sepsis	150	0	83
Data source	Xuzhou Central Hospital/Northern Theater General Hospital/ Eastern Theater General Hospital	Northern Theater General Hospital	Eastern Theater General Hospital

Fig. 1.

The graph of the study process.

The CT images were taken while the patients were lying down with their heads raised, using a slice thickness ranging from 1 to 5 mm. Two CT scanners, the Philips Brilliance 64-layer scanner and the Siemens SOMATOM Definition Flash CT scanner, were used to capture the images. The tube voltage ranged from 80 to 140 kilovolts, and the tube current ranged from 43 to 1410 milliamperes. Both plain and contrast-enhanced CT images were used.

Ground-truth labeling and area calculation

Labeling L3 CT in training and validating datasets is carried out by two radiologists with senior professional titles. They label three types of body compositions, namely skeletal muscle (SM), visceral adipose tissue (VAT), and subcutaneous adipose tissue (SAT), with the help of 3D-slicer software. The same label procedure was performed on Cohort 2 and Cohort 3.

To calculate the skeletal muscle area (SMA), Hounsfield unit (HU) thresholds of − 29 and + 150 are used. The subcutaneous adipose area (SATA) is calculated from extra muscular tissue with a density between − 190 and − 30 HU, and the visceral adipose tissue area (VATA) is calculated from non-subcutaneous tissue with a density between − 150 and − 50 HU. The calculated result was used for the consistency test. All of the areas were measured in cm² and normalized by the patient’s height (m) to calculate the skeletal muscle index (SMI), subcutaneous adipose tissue index (SATI), and visceral adipose tissue index (VATI).

Data preprocessing

Before data model training, several image transforms were applied to improve the training efficacy and robustance of the model, including data normalization, pixel normalization, pixel scaling, random 90-degree rotation, random inversion, and random Gaussian noise. Additionally, data augmentation was performed to the available training set, primarily random rotation, horizontal flipping, and vertical flipping, to reduce the possibility of framework overfitting and enhance generalization.

The construction of the L3BCSM

The UNet network is used as the basic framework of the model. The structure of the UNet network consists of an encoder (down-sampling path) and a decoder (up-sampling path). The encoder comprises five convolutional layers, with a down-sampling operation following each convolutional layer. This gradually reduces the feature map’s size while increasing the feature’s abstraction level. The decoder also consists of five convolutional layers, each followed by an up-sampling operation combined with a feature map from the attention gate. This gradually increases the size of the feature map while restoring the spatial resolution.

In the L3BCSM, the decoder stage of UNet consists of five layers, with the first four introducing the attention gate and Squeeze and Excitation block (SEblock). The attention gate captures spatial information, while the SEblock captures channel information. The fifth layer includes an Atrous Spatial Pyramid Pooling (ASPP) that captures scale information. These three blocks work together to provide spatial, channel, and scale attention, which helps extract essential features and improve the segmentation performance of the model (Fig. 2).

Fig. 2.

The graph of L3BCSM structure. The UNet network is used as the basic framework of the model, including an encoder (down-sampling path) composed of five convolutional layers and a decoder (up-sampling path) consisting of another five convolutional layers. The first four layers of UNet’s decoder stage introduce the attention gate, Squeeze, and Excitation block (SEblock). An attention gate is adopted to capture spatial information, while SEblock is adopted to capture channel information. The fifth decoder layer comprises an Atrous Spatial Pyramid Pooling (ASPP), which can capture scale information.

Training

Three models, L3BCSM, UNETR, and AHNet, were used with the same parameters and training dataset during the training process. The Loss Function used was Dicefocalloss, and the Optimization Method was Adam. The initial learning rate was set to 0.001, and the learning rate of the optimizer was reduced by 2% every epoch. The other hyperparameters include batch size of 2, input image size of 512*512 pixels, and training epochs of 100. After each epoch training, the weight file .ptw was saved, and the best one was selected. The training was performed on Nvidia GeForce RTX 3080 using version 3.9.6 of the Python programming language, published by the Python Software Foundation headquartered in Beaverton, Oregon, USA. PyTorch 1.9.0 was used as a deep learning framework for the training process.

Internal validation and external test of models

Several metrics were used for internal validation and external tests, including the Dice similarity coefficient (DSC), Jaccard coefficient, HD95 (the 95th percentile of the Hausdorff distance), precision, and recall. During the internal validation, the performance of three models was observed. However, only the L3BCSM model was used in the external tests conducted on two test datasets. In comparing the performance of models, the Mann–Whitney U test was performed between L3BCSM and the other two models. A P < 0.05 was considered statistically significant. Before conducting the statistical analysis, the Shapiro–Wilk test was used to assess the data distribution. The results indicated that all data followed a non-normal distribution, leading to the selection of the Mann–Whitney U test.

Consistency test and manual result checking

Bland–Altman plots of each body composition, including SM, SAT, and VAT, are used to reflect the consistency between L3BCSM model labeling and manual labeling on Cohort 2 and 3 datasets (Fig. 3). Also, the segmentation results of each body composition, including SM, SAT, and VAT, were manually checked on the randomly selected images from Cohort 2 and 3 datasets.

Fig3.

The Bland–Altman plots of each composition in Cohort 2 and 3. Skeleton muscle area (SMA) (a), subcutaneous adipose area (SATA) (b), and visceral adipose tissue area (VATA)(c) in Cohort 2; Skeleton muscle area (SMA) (d), subcutaneous adipose area (SATA)(e), and visceral adipose tissue area (VATA)(f) in Cohort 3.

GUI design and online deployment

The application consists of five modules: the “About” module, the “Population Analysis” module, the "Time Series Analysis" module, the "Manual Result Checking" module, and the “Consistency Testing” module. The “About” module provides an introduction to the background and usage method and a description of the model’s efficiency. The other modules perform specific functions.

When the user uploads CT images in batches, the application automatically segments the body compositions on the CT image, including SM, SAT, and VAT. The application requires additional parameters to calculate the SMI, SATI, and VATI, which can be easily inputted through the provided interface. The results are displayed through images, tables, and charts, which can be downloaded for further analysis.

The application is built using the Python package Streamlit, which allows for easy construction of the GUI and deployment on the Streamlit Cloud for framework access.

Results

Comparing the performance of L3BCSM with the UNETR and AHNet models

The effectiveness of different models, such as SAT, VAT, and SM, in assessing body composition on CT images was evaluated using a comprehensive set of metrics, including Dice coefficient, Jaccard index, Hausdorff distance (HD95), precision, and recall on the internal validation dataset. The L3BCSM model achieved satisfactory scores comparable to the UNETR and AHNet models on all metrics. Specifically, L3BCSM performed better in all metrics of SATA and SMA than UNETR (all P < 0.05) and was comparable in VATA. While the performance of L3BCSM was comparable to AHNet in VATA and SMA and better in SATA (most P < 0.05, except P of HD95) (Table 2).

Table 2.

Internal evaluation and comparison of the three models (median (IQR)).

Models	Composition	Dice	Jaccard	HD95	precision	Recall
L3BCSM	SATA	0.954(0.930, 0.963) P = -	0.912(0.869, 0.928) P = -	3.350(2.20, 5.33) P = -	0.968(0.954,0.981) P = -	0.944(0.910,0.9625) P = -
	VATA	0.849(0.774,0.901) P = -	0.737(0.632,0.820) P = -	17.0(10, 29) P = -	0.897(0.783,0.938) P = -	0.863(0.768,0.928) P = -
	SMA	0.920(0.901, 0.936) P = -	0.852(0.820,0.880) P = -	7.800(5.6,10) P = -	0.934(0.905,0.964) P = -	0.913(0.881,0.938) P = -
UNETR	SATA	0.927(0.892,0.948) P < 0.001	0.894(0.841,0.912) P < 0.001	9.400(5.00,21.44) P < 0.001	0.941(0.899,0.971) P < 0.001	0.921(0.875,0.949) P < 0.001
	VATA	0.829(0.749,0.875) P = 0.110	0.722(0.580,0.806) P = 0.110	24.50(13.00,42.00) P = 0.0064	0.838(0.729,0.923) P = 0.038	0.845(0.769,0.910) P = 0.321
	SMA	0.856(0.801,0.888) P < 0.001	0.852(0.817,0.879) P < 0.001	39.00(23.00,49.00) P < 0.001	0.857(0.805, 0.910) P < 0.001	0.862(0.796,0.897) P < 0.001
AHNet	SATA	0.944(0.913,0.955) P = 0.002	0.864(0.805,0.902) P = 0.002	3.000(2.000,4.130) P = 0.454	0.961(0.933,0.976) P = 0.010	0.926(0.885,0.955) P = 0.048
	VATA	0.839(0.735,0.892) P = 0.075	0.708(0.598,0.778) P = 0.755	20.00(10.75,35.5) P = 0.154	0.861(0.771,0.899) P = 0.0063	0.847(0.738,0.922) P = 0.418
	SMA	0.920(0.899,0.936) P = 0.458	0.746(0.668,0.798) P = 0.456	7.050(5.00,10.25) P = 0.669	0.933(0.897,0.957) P = 0.377	0.917(0.871,0.941) P = 0.946
The Mann–Whitney U test was performed between L3BCSM and the other two models

External evaluation of the L3BCSM model on the test datasets

L3BCSM showed satisfactory performance in model efficacy metrics in Cohort 2, as well as in Cohort 3. For instance, the median Dice score was 0.919 (0.842,0.943) in SATA, 0.741(0.648, 0.759) in VATA, and 0.858(0.798,0.876) in SMA in Cohort 2, whereas in Cohort 3, the median Dice score was 0.927(0.863,0.969) in SATA, 0.803(0.678,0.893) in VATA, and 0.880 (0.863,0.901) in SMA. All other metrics were consistent with the median Dice score (Table 3).

Table 3.

External evaluation of L3BCSM (median (IQR))

Cohort	Composition	Dice	Jaccard	HD95	precision	Recall
Cohort 2	SATA	0.919(0.842,0.943)	0.852(0.726,0.892)	4.00(2.20,10.75)	0.917(0.869,0.949)	0.932(0.841,0.957)
	VATA	0.741(0.648, 0.7759)	0.472(0.289,0.634)	57.5(34,67.25)	0.548(0.423,0.683)	0.810(0.632,0.898)
	SMA	0.858(0.798,0.876)	0.751(0.663,0.779)	7.8(5.525,11.25)	0.899(0.842,0.935)	0.811(0.757,0.847)
Cohort 3	SATA	0.927(0.863,0.969)	0.867(0.759,0.939)	4.3(3.075,7.525)	0.977(0.964,0.988)	0.880(0.782,0.957)
	VATA	0.803(0.678,0.893)	0.671(0.513,0.806)	32.5(22.25,43.00)	0.823(0.749,0.910)	0.824(0.621,0.897)
	SMA	0.880(0.863,0.901)	0.785(0.759,0.820)	8.10(5.60,11.75)	0.913(0.863,0.961)	0.850(0.813,0.917)

Consistency test and manual result checking

The area of each composition was calculated independently by both L3BCSM and humans. In both Cohort 2 and Cohort 3, the difference point was within the 95% Limits of Agreement in the Bland–Altman plot. All three body compositions had a P value greater than 0.05 in their respective student’s t-tests (Fig. 3). Additionally, three sets of CT images and their L3BCSM labeling and human labeling in Cohort 2 and 3 were randomly selected to check the segmentation results manually. In Cohort 2 and Cohort 3, only a slight difference was found between L3BCSM labeling and human labeling, and both were qualified labeling (Fig. 4).

Fig4.

Manually checking the segmentation results of L3BCSM and humans in Cohorts 2 and 3. The original image, prediction, and label of cases in Cohort 2(a), and the magnified original image, prediction, and label of cases in Cohort 2 (b). The original image, prediction, and label of cases in Cohort 3(c), and the magnified original image, prediction, and label of cases in Cohort 3(d). The white dot squares indicate the contents which were magnified.

GUI for multiple application scenes was deployed

This application has four functional modules except the “About” module. These modules can be classified into two categories: analysis and supporting modules. The “Population analysis” and "Time series analysis" modules belong to the analysis modules, while the “Consistency testing” and "Manual result checking" modules are the supporting modules.

If the user wants to acknowledge the distribution of body compositions of a specific population, they can upload the CT image to the “Population analysis” module. This will provide them with the body composition distribution in that population. On the other hand, if the user wants to observe the changes in body composition of the same patient at different time points, they can upload the CT image to the "Time series analysis" module. This will provide them with the calculation result of body compositions, and they can enter the corresponding time point to display the calculation result using a line chart.

Before using the application, it is necessary to determine whether the new dataset is suitable for using the application. The supporting modules were designed to achieve this goal. The user can perform a consistency test in the “Consistency testing” module and manually check the segmented effect by comparing the difference between the model predicting mask and the manual label in the "Manual result checking" module (Fig. 5). The application is accessible at https://body-compositions-assessment-tool.streamlit.app/.

Fig5.

GUI design of body composition assessment application. The application is accessible at https://body-compositions-assessment-tool.streamlit.app/.

Discussion

The in vivo evaluation of body composition is essential in many clinical investigations to accurately describe and monitor the status of a range of medical conditions in patients with cancer^19,20, osteoporosis²¹, and many other diseases²². In most research and clinical practices, assessing body compositions in CT images, such as skeletal muscle, visceral adipose, and subcutaneous adipose, requires manual labeling and measurement. This repetitive and tedious process significantly limits related research and clinical practice progress. To overcome this challenge, we have developed a web application that uses a well-performed model to automatically identify skeletal muscle and adipose tissue in L3 CT images and calculate their area. This highly automated approach opens up the possibility of large-scale screening, making it easier for doctors to estimate body composition and use it for research and diagnosis. For example, precise body composition measurement can help determine insulin dosage and dietary advice²³; managing the body composition of cancer patients may improve their prognosis²⁴.

In our study, L3 level CT image was chosen for body composition measurement, regarded as the gold standard level for body composition assessment²⁵. The L3 level includes essential muscle groups and fat distributions, including the psoas major, rectus abdominis, latissimus dorsi, and subcutaneous and visceral fat. The muscle and adipose mass at the L3 level correlates highly with the whole body^25,26. Numerous studies used the L3 CT image to assess the body compositions^7,27 , making it easier to compare and validate results with each other. In addition to utilizing L3 CT images, another approach to assess body composition involves using multi-level CT images^9,10. This method utilizes a 3D encoder and decoder in its network to precisely measure body composition by handling more data and requiring more computing power²⁸. This approach would be an excellent option for accurately measuring body composition, especially in cases where the L3 CT image is unavailable²⁸.

Regarding model construction, we experimented with different attention algorithms to improve the segmentation performance of the Unit framework. We incorporated attention gates, SEblock, and ASPP into every model layer. This resulted in the L3BCSM model being as effective as other segmentation frameworks like UNETR²⁹ and AHNet³⁰. The Attention UNet architecture uses attention gates to enhance feature representations during the upsampling and downsampling processes. The attention gate is a crucial component that connects the encoder and decoder paths in a neural network. It comprises two main branches: the Skip Connection and the Complementary Path³¹. The Skip Connection copies the high-resolution feature map from the encoder path and passes it to the decoder path. The Complementary Path is a low-resolution feature map from the encoder path that undergoes a set of operations, typically involving convolutional and activation layers, to compute attention weights adaptively. These attention weights represent the importance of each spatial location in the low-resolution feature map concerning the corresponding locations in the high-resolution feature map³². The attention weights are then applied to the Skip Connection, emphasizing or de-emphasizing different spatial locations based on their importance. This adaptive fusion of features helps the network focus on relevant details during upsampling. The SEblock structure enhances convolutional neural networks (CNNs) by capturing channel-wise dependencies. It involves two steps: squeeze and excitation. The squeeze step reduces spatial dimensions, yielding a 1D vector for each channel via global average pooling. The excitation uses fully connected layers to capture inter-channel dependencies and produce scaling factors. The final output is obtained by scaling the original feature maps with the computed factors, allowing the network to emphasize or suppress specific channels adaptively for improved representation³³. The ASPP structure, introduced in DeepLab, consists of parallel atrous convolutional layers that capture multi-scale information and a global average pooling layer for global context³⁴. These features are then combined and processed to generate refined predictions in tasks like semantic segmentation. This design enables the network to effectively understand local and global contextual information, leading to superior performance. The attention mechanism is widely used to improve the efficiency of the segmentation model. Different types of attention mechanisms can be combined to enhance model performance. For instance, the Unet model incorporates Attention Gating Modules and a Squeeze and Excitation block (SEblock) to focus on regions of interest and improve feature map quality¹³. Similarly, a novel dense residual UNet model combines AG, efficient channel attention (ECA), and ASPP to enhance the performance of the automatic multiple sclerosis lesion segmentation using 3D MRI sequences was also constructed¹⁴. We integrated three types of attention mechanisms – AG, SEBlock, and ASSP – into a Unet-based model, which outperformed UnetR and Ahmet. According to our knowledge, we are the second group to integrate three types of attention mechanisms into a single model and the pioneers in utilizing this approach in CT images.

Using attention gates, SEblock, and ASPP in the UNet architecture allows the model to selectively attend to informative regions and channel and scale level information, improving its ability to capture fine details and handle complex relationships between features in medical image segmentation and other tasks. The L3BCSM is a body composition assessment model using multiple CT image sources as the training dataset, including liver, gastric, and colorectal cancer patients, et al. This makes it more robust in clinical practice in gastroenterology compared with existing models in which a single image source was used as a training dataset^7,19,35. The external test results in Cohort 2 and Corhort 3 confirm this. Meanwhile, the training dataset contains two types of CT images: plain scan and enhanced-contrast scan. Therefore, the model was supposed to adapt to both kinds of CT images. Based on these characteristics of training data, the ideal parameters for the input CT image for L3BCSM are L3 CT images with plain or enhanced-contrast scans from patients with gastroenterological tumors or abdominal infections.

To assess the generalizability and robustness of L3BCSM, multiple external tests were conducted using CT images from different medical centers. These tests utilized datasets with diverse parameters, allowing for a comprehensive evaluation of the model’s performance. Cohort 2 and Cohort 3 consisted of a dataset with parameters similar to the L3BCSM training dataset, including CT image types (plain and enhanced-contrast) and patient demographics (abdominal disease patients). As expected, L3BCSM performed well on this dataset, demonstrating its effectiveness for body composition segmentation within its intended domain.

The L3BCSM model serves as the central component of this application, driving its functionality and enabling various usage scenarios. It seamlessly interacts with other complementary elements, including data frame creation, plotting, and statistical analysis, to present a comprehensive and informative user experience. The application offers four distinct functionalities, each catering to specific needs. The “Population Analysis” module allows users to explore and understand the distribution of body compositions within specific populations, which is the most common study type^7,36. L3BCSM’s output is presented as data tables and visualizations, providing insight into the data characteristics. The “Time Series Analysis” module provides a platform for tracking individual patients’ temporal changes in body composition over a defined period, like in a case study following a drug-free male bodybuilder for six months³⁷. L3BCSM’s output is displayed as data tables and charts, enabling users to monitor and analyze the body composition trends within a patient. Before utilizing the L3BCSM model, it’s necessary to ensure that the input CT image meets the model’s requirements and that the calculated results are reliable. In the “Consistency Test” module, Bland–Altman plots and student’s t-tests are used to reflect the consistency between methods³⁸, which were used as supporting modules to assess the consistency and accuracy of the model’s predictions. The “Manual Result Checking” module allows users to manually verify the L3BCSM’s segmentation results, providing more confidence in the model’s performance besides the above consistency test. This is particularly helpful when working with new datasets and indicating which method is the better. Before using the “Population Analysis” module, it is recommended that the “Consistency Test” and "Manual Result Checking" modules be run on the new dataset. This ensures the application’s compatibility with the new dataset and facilitates reliable analysis. As a practice example, one of our collaborators in another hospital designed a clinical trial to reveal the predicting effect of body compositions on the prognosis of gastric cancer patients receiving PD-1 inhibitors-based combination therapy³⁹. The researchers gathered over one hundred CT images to analyze the body composition parameters to achieve this goal. They used this app to assist in extracting these parameters. They initially selected 1/3 of the CT images and manually labeled them for the three types of body compositions. They then utilized the “Consistency Test” and "Manual Result Checking" modules to ensure the app was suitable for their data. After confirming its suitability, they used the “Population analysis” modules to extract the body composition parameters from all the CT images.

The current L3BCSM model may work best with data similar to the Gastrointestinal Surgery patient dataset it was trained on and may not perform well in extreme conditions. To improve this, we should consider using a more diverse training dataset in the future. Generally, increasing the training data size would lead to better model performance and is one way to enhance the L3BCSM. Training the model using jointed CT images from multiple sites could be an effective way to enhance its performance, as recent research indicates¹¹. Additionally, 3D segmentation using either 3D or 2D U-Net has shown significant potential, demonstrating improved segmentation accuracy compared to current methods^9,11. We believe that this technique represents a next-generation approach for segmenting and evaluating body compositions. Besides, although several commonly used functions have been integrated into the application, they have not been thoroughly tested in practice. Therefore, we will modify the app’s functions based on their performance in real-world scenarios.

Conclusion

In this study, we developed a high-performing deep learning model called L3BCSM and integrated it into multiple usage scenes, providing a convenient tool for sarcopenia research and screening.

Author contributions

Liang Zhang was responsible for designing the study, collecting the CT images and preparing the primary manuscript. Jiao Li assisted in collecting the CT images, preparing the manuscript and coding the APP. Zhi Yang and Lin Zhang designed and coded the model. Jun Yan and Long-bo Gong provided direction for the study design and edited the final manuscript. All authors reviewed the manuscript.

Data availability

The data presented in this work are not publicly accessible due to their enormous size but are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.