Segmentation and diagnosis of anterior cruciate ligament tear using deep learning and radiomics based on knee CT

Abstract

Purpose

Timely and accurate diagnosis of anterior cruciate ligament (ACL) tears has important clinical significance. In this study we tried to establish a segmentation and diagnosis model for ACL tear using deep learning and radiomics based on knee CT.

Materials and methods

Totally 469 patients were collected for ACL segmentation model construction. Among them, 328 patients underwent MRI examination within one week of CT scanning and were used to construct diagnosis model. The segmentation model was trained using deep learning of 3D nnU-Net. After segmentation, a total of 2,264 quantitative radiomics features were extracted from each ACL. The support vector machine (SVM), random forest (RF) and stochastic gradient descent (SGD) were used to construct classification model.

Results

The 3D nnU-Net segmentation model we constructed achieved high performance in the ACL segmentation with Dice Similarity Coefficient (DSC) of 0.79 in the external validation. In terms of ACL tear diagnosis, the SVM, RF, and SGD models all demonstrated excellent performance. In the external validation, the Area Under the Curve (AUC) were 0.85, 0.86, and 0.81.

Conclusions

We developed a CT based artificial intelligence system that could perform ACL segmentation and tears diagnosis. It had high accuracy and convenience, and was of great significance in clinical practice.

Introduction

The anterior cruciate ligament (ACL) plays a crucial role in maintaining knee joint stability and normal motor function [1]. Tearing of ACL can cause instability of the knee joint, symptoms such as joint pain and limited mobility, and lead to serious complications including knee osteoarthritis, meniscus and cartilage damage [2]. According to the clinical diagnosis and treatment guidelines for ACL injury [3], timely and accurate diagnosis of ACL tear has important significance in clinical practice.

Arthroscopy is the “gold standard” for ACL tears diagnosis [4], but it is an invasive examination method which has surgical risks and potential complications. Currently, magnetic resonance imaging (MRI) is the preferred imaging modality for ACL tears examination [5]. However, the long examination time and high cost limited its clinical application. Especially for individuals with claustrophobia and those who have metal devices such as pacemakers and insulin pumps in their bodies, MRI is contraindicated. Therefore, the use of computed tomography (CT) for ACL tears diagnosis is necessary, especially for patients who cannot undergo magnetic resonance imaging. CT has the advantages of simple operation, fast speed, wide application, low examination cost, and is the preferred emergency examination method for bone and joint injuries. However, in clinical practice, reconstructing and delineating ACL based on CT images are very time-consuming, and due to the low soft tissue resolution of CT, the accuracy in diagnosing ACL tears is very low, far from reaching clinical application. In recent years, artificial intelligence of deep learning have rapidly developed in the field of medical imaging, and can be used for medical image reconstruction and segmentation [6]. Radiomics [7, 8] is a high-throughput, automated imaging data analysis method that, when combined with machine learning, has shown great potential for accurate diagnosis and classification of bone and joint diseases including tumors, inflammation, and degeneration. In previous studies [9–11], deep learning, machine learning, and radiomics have been used for ACL segmentation and tear diagnosis in MRI images, demonstrating high accuracy. However, there is still few report on ACL segmentation and diagnosis based on CT images. Therefore, the purpose of this study was to establish an ACL segmentation model using deep learning, and then use machine learning to construct a radiological model to achieve CT diagnosis of ACL tears, helping with early treatment in clinical practice.

Materials and methods

Subjects

This study was approved by the Ethics Committee of Chongqing University Central Hospital (Hospital 1) and Chongqing University Fuling Hospital (Hospital 2). As a retrospective study, patient informed consent was waved. A total of 525 patients from June 2016 to February 2024 in the First Hospital were collected. According to the exclusion criteria, 469 patients were finally included for the construction of ACL segmentation model. The inclusion criteria were as follows: (1) Age ≥ 18 years old; (2) Clinical indications of acute or chronic pain and trauma; (3) Complete CT images of the knee joint; (4) slice thickness of CT images was less than 1 mm. The exclusion criteria were: (1) Patients who had undergone ACL surgery in the past; (2) CT images with severe artifacts. Among them, 328 patients underwent MRI examination within one week of CT scanning and were divided into ACL tear group and intact ACL group based on MRI diagnosis, which were used to construct diagnosis model. According to the same inclusion and exclusion criteria, 85 patients were collected for external validation of the segmentation model, 70 patients were used for external validation of the classification mode from Hospital 2 between November 2021 and October 2024. Figure 1 showed the patient enrollment process and study flow chart.

Fig. 1.

Flowchart of patients enrollment and study design

CT examination and image preprocessing

All CT images were obtained from 7 different CT scans: General Electric Company, Optima CT 660, USA; General Electric Company, Revolution CT 256, USA; General Electric Company, Light Speed VCT 64, USA; Siemens Company, SOMATOM Definition Flash, Germany; Philips Company, Brilliance iCT, the Netherlands; Siemens Company, SOMATOM Force Dual Source CT Scanner, Germany; and United Imaging uCT 760, Shanghai, China. Similar scanning parameters were used for all scans: tube voltage of 100 to 120 kV, automatic mas; matrix size 512 × 512; layer thickness: 0.5 to 0.625 millimeters; reconstruction thickness: 1 mm; reconstruction interval: 0.5 to 0.625 mm. All digital imaging and communications in medicine (DICOM) format image data were preprocessed using the uAI Research Platform (version 20241130) [12]. The reconstruction and annotation of ACL were performed by two radiologist with over 10 years of experience, serving as the gold standard for segmentation models training. The preprocessing process included: (1) spatial standardization: eliminating spatial resolution differences between devices through voxel resampling of 1 mm × 1 mm × 1 mm, and applying standardized display parameters with a window width of 300 Hounsfield Unit (HU) and a window level of 50 HU to enhance the contrast of ACL areas; (2) grayscale standardization: truncate CT values to -1,000 ~ 1,000 HU and normalize them to the 0–1 range to eliminate differences in grayscale response among devices.

Deep learning-based ACL segmentation

The ACL segmentation model was constructed based on 469 patients at Hospital 1. Random stratified sampling was used to divide them into a training set (375 cases) and an internal testing set (94 cases). Subsequently, 85 patients collected from Hospital 2 were used as an independent external validation set to test the generalization ability and robustness of the model. The segmentation model training was performed using standard 3D nnU-Net framework (version 1.7.0) (https://github.com/MIC-DKFZ/nnUNet) [13], which automatically determined the optimal network configuration and training parameters (such as batch size, learning rate, data augmentation scheme) based on the provided training data based on PyTorch 1.10.0 and 4 x NVIDIA RTX A4000 Graphics Processing Unit (GPU) (Ubuntu 20.04 system). On the basis of inheriting the multi-scale feature extraction capability of classical U-Net, the 3D nnU-Net achieved end-to-end optimization through adaptive parameter configuration (including network depth, convolution kernel size, and data enhancement strategy), avoiding the high dependence of traditional methods on manual parameter tuning. It supported 2D/3D mixed training, could fully explore the three-dimensional spatial correlation of CT images, and accurately modeled the complex anatomical morphology and individual variation characteristics of ACL. It had a sliding window prediction algorithm and a threshold based post-processing module, significantly reducing GPU memory consumption and adapting to the efficient processing requirements of large medical images. During the training process, data augmentation was applied on-the-fly and exclusively to the training set to improve generalization. This included random rotation (± 15 °), scaling (0.85–1.15 times), and Gaussian noise injection (σ = 0.05) were applied for data augmentation, while mirror flipping was excluded to avoid left-right confusion. In terms of parameter optimization, the framework automatically determined the optimal batch size (batch size = 4) and initial learning rate (0.0003), using the Adam optimizer and Dice Cross Entropy. The models were trained through 1000 iterations using an early stopping mechanism (patch = 50) to prevent overfitting. During the training process, the 3D nnU-Net automatically performed five fold cross validation internally, monitoring model performance using a validation set and selecting the optimal model. The final trained model first undergo preliminary performance evaluation on the internal test set, and then its generalization ability and robustness were tested on the external test set. The performance of the model was quantified through three dimensions: geometric consistency, boundary accuracy, and clinical applicability. Four indicators including Dice Similarity Coefficient (DSC), Relative Volume Error (RVE), 95% Hausdorff Distance (HD95), and Average Symmetric Surface Distance (ASSD) were used for quantitative analysis. DSC and RVE focused on overall volume consistency, while HD and ASSD focused on local boundary errors.

Radiomics -based ACL tears diagnosis

Totally 328 patients were divided into ACL tear group and intact ACL group based on MRI diagnosis. The diagnosis of ACL tear were made by two senior radiologists with more than 15 years experience. All of them were randomly divided into a training set (116 patients with ACL tears and 146 patients with intact ACL) and a testing set (30 patients with ACL tears and 36 patients with intact ACL) in an 8:2 ratio. Another 38 patients with ACL tear and 32 patients with intact tear from Hospital 2 were used for external validation.

All radiomics features were extracted from the segmented ACL by 3D nnU-Net using PyRadiomics (version 3.0.1). Feature extraction was performed using PyRadiomics (v3.0.1) and Scikit-learn (v0.23.2). The sequence and scope of steps were: image resampling → feature extraction → Z-score normalization → LASSO feature selection(only in the training set) → MaxAbsScaler scaling → LDA dimensionality reduction → classifier training → independent test set evaluation. The Image resampling included target voxel spacing (1 × 1 × 1 mm³) and interpolation method (cubic B-splines). Box width was fixed with 25 HU for intensity discretization. A total of 2,264 quantitative radiomics features were extracted from each ACL, including shape features, grayscale statistical features, texture features, and high-order features. Among them, the shape features included 14 three-dimensional geometric attributes such as maximum diameter, volume, and surface area. The grayscale statistics characteristics included 18 indicators that reflected voxel intensity distribution, such as mean, peak, and maximum. Texture features were calculated based on 72 spatial distribution attributes, including gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), and gray level region matrix (GLSZM). High order features were obtained through 24 types of filters, including Box Mean, additive Gaussian noise, binomial blur, curvature flow, etc. Box-Sigma, Standardization, Laplacian sharpening, discrete Gaussian, mean, speckle noise, recursive Gaussian, shot noise, and wavelet transform LLL, LLH, LHL, LHH, HLL, HLH, HHL, and HHH. The least absolute shrinkage and selection operator (LASSO) was used to gradually screen for the most robust radiomics features. All preprocessing, feature selection and dimensionality reduction steps fit parameters on the training set, and then transform the test set with the same parameters to strictly prevent data leakage.

Three classifiers of SVM with RBF kernel, Random Forest (RF, 100 trees), and Stochastic Gradient Descent (SGD) with hinge loss were trained on the training set (Scikit-learn, v 0.23.2). The performance of the models were evaluated using the Area Under the Curve (AUC), Sensitivity, Specification, Accuracy, and F1 score. Delong test was used for the comparison of different models. A calibration curve was conducted to evaluate the consistency (goodness of fit) of the model between predicted and actual results. The decision curve analysis (DCA) was performed to quantify the clinical utility value. The segmentation and classification models construction process was shown in Fig. 2.

Fig. 2.

The workflow diagram of ACL segmentation model construction (a); The workflow diagram of Radiomics analysis and models establishment for ACL tears diagnosis (b)

Statistic analysis

All statistical analyses were conducted using Python 3.7 software. The quantitative data that conformed to normal distribution was expressed as mean ± standard deviation, and the group comparison was conducted using two independent sample t-tests. Quantitative data that did not conform to normal distribution were represented by median and interquartile range [M (P25, P75)], and the Mann Whitney U test was used for group comparison. The qualitative data were described as frequency (percentage), and the groups comparison was performed using a chi square test. DeLong test were used to compare the performance among different models. The P-value < 0.05 was considered statistically significant in all analyses.

Results

Patient characteristics

This study included 469 patients in Hospital 1 as the training and testing sets for 3D nnU-Net segmentation model establishment. Among them, there were 227 males (60.5%) and 148 females (39.5%) in the training set, with an average age of 50.0 years. There were 53 males (56.4%) and 41 females (43.6%) in the testing set, with an average age of 48.5 years. A total of 85 patients were included as the independent external test set from the Hospital 2, including 50 males (58.8%) and 35 females (41.2%), with an average age of 50.0 years. There was no statistical difference in age (p = 0.597) and gender (p = 0.756) among the three groups. For the construction of diagnostic models, a total of 328 patients were included from Hospital 1 as the training and testing sets. Among them, there were 149 males (56.9%) and 113 females (43.1%) in the training set, with an average age of 49.0 years. There were 34 males (51.5%) and 32 females (48.5%) in the test set, with an average age of 50.0 years. 70 patients were included from Hospital 2 as an independent external test set, with 46 males (65.7%) and 24 females (34.3%), with an average age of 50.0 years. There was no statistical difference in age (p = 0.543) and gender (p = 0.230) among the three groups (Table 1).

Table 1.

Dmeographic characteristics of patients in this study

	Characteristics		Training cohort	Testing cohort	External Validation cohort	p-value
ACL Segmentation Model Construction	Age (Median [Q1,Q3])		50.000 [36.500,60.500]	48.500 [31.250,61.500]	50.000 [37.000,60.000]	0.597
	Gender (%)					0.756
		Male	227 (60.533)	53 (56.383)	50 (58.824)
		Female	148 (39.467)	41 (43.617)	35 (41.176)
ACL Tears Diagnosis Model Construction	Age (Median [Q1,Q3])		49.000 [33.000,61.000]	50.000 [36.000,62.000]	50.000 [44.250,60.000]	0.543
	gender (%)					0.230
		Male	149 (56.870)	34 (51.515)	46 (65.714)
		Female	113 (43.130)	32 (48.485)	24 (34.286)

Segmentation model performance

The 3D nnU-Net we constructed achieved high performance in the ACL segmentation. On the training set, the DSC was 0.85 ± 0.02, the RVE was 0.08 ± 0.07, the HD95 was 3.44 ± 1.42 mm, and the ASSD was 0.96 ± 0.20 mm. On the test set, the DSC was 0.81 ± 0.04, the RVE was 0.16 ± 0.13, the HD95 was 12.00 ± 67.47 mm and the ASSD was 1.48 ± 1.76 mm, respectively. On the independent external validation set, the DSC was 0.79 ± 0.04, the RVE was 0.30 ± 0.19, the HD95 was 5.52 ± 2.40 mm and ASSD was 1.39 ± 0.40 mm, respectively. The segmentation results of the models were shown in (Table 2).

Table 2.

Results of the segmentation model

	Dice	RVE	HD95(mm)	ASSD(mm)
Training set (n = 375)	0.85 ± 0.02	0.08 ± 0.07	3.44 ± 1.42	0.96 ± 0.20
Testing set (n = 94)	0.81 ± 0.04	0.16 ± 0.13	12.00 ± 67.47	1.48 ± 1.76
Independent external validation set (n = 85)	0.79 ± 0.04	0.30 ± 0.19	5.52 ± 2.40	1.39 ± 0.40

The Dice coefficient measures the degree of overlap between the predicted segmentation result and the ground truth label, and it is used to evaluate the segmentation accuracy. Its range is from 0 to 1, where 1 indicates a complete overlap and 0 indicates no overlap. The Hausdorff distance measures the maximum distance between two point sets and is used to assess the accuracy of the segmentation boundary. A smaller value indicates that the segmentation boundary is closer to the true boundary. The average symmetric surface distance measures the average symmetric distance between two surfaces. A smaller value indicates that the two surfaces are closer, and it is used to evaluate the accuracy of the segmentation surface

Diagnosis model performance

In terms of ACL tear diagnosis, after feature screening by LASSO, there were 78 features left for modeling, including 18 strength features, 3 shape features, 44 texture features, and 13 high-order features. The SVM, RF, and SGD models all demonstrated excellent performance. In the training set, the AUC of SVM, RF, and SGD models were 0.97 (95% CI: 0.95–0.98), 0.97 (95% CI: 0.95–0.98), and 0.93 (95% CI: 0.89–0.96), respectively. In the test set, their AUC were 0.92 (95% CI: 0.84–0.97), 0.92 (95% CI: 0.84–0.97), and 0.89 (95% CI: 0.81–0.96), respectively. In the independent external validation set, their AUC were 0.85 (95% CI: 0.75–0.94), 0.86 (95% CI: 0.77–0.94), and 0.81 (95% CI: 0.72–0.90), respectively (Table 3). The DeLong test showed that there was no significant difference among these three models (P = 0.099,P = 0.595,P = 0.124). The calibration plots showed that there was a good consistency between the prediction and true results in the training cohort and the test cohort. The decision curve plot showed that the model achieved great benefit and clinical significance for patients with ACL tears (Figs. 3 and 4).

Table 3.

Performance of the three classification models for ACL tears diagnosis

Models	Group	AUC	sensitivity	specificity	accuracy	f1-score
SVM	Train cohort	0.97(0.95–0.98)	0.88	0.92	0.90	0.89
	Test cohort	0.92(0.84–0.97)	0.83	0.83	0.83	0.82
	Validation cohort	0.85(0.75–0.94)	0.74	0.84	0.79	0.80
RF	Train cohort	0.97(0.95–0.98)	0.88	0.93	0.91	0.90
	Test cohort	0.92(0.84–0.97)	0.80	0.83	0.82	0.80
	Validation cohort	0.86(0.77–0.94)	0.72	0.88	0.79	0.79
SGD	Train cohort	0.93(0.89–0.96)	0.92	0.85	0.88	0.88
	Test cohort	0.89(0.81–0.96)	0.93	0.78	0.85	0.85
	Validation cohort	0.81(0.72–0.90)	0.77	0.81	0.79	0.80

SVM: Support Vector Machine RF: Random Forest SGD: Stochastic Gradient Descent

Fig. 3.

The radar plots were used to display the diagnostic performance of the three models (SVM, RF, and SGD) in the training set, testing set and external testing set

Fig. 4.

The Receiver Operating Characteristic (ROC) curves of SVM, RF, and SGD models in training set, testing set and external testing set (a-c). The calibration curves of the SVM, RF, and SGD models in the training set, testing set and external testing set. The 45-degree line represents the scenario where the model-predicted probabilities match the actual probabilities; the closer the two curves are, the higher the accuracy (d-f). The decision curves of the SVM, RF, and SGD models in the training set, testing set and external testing set, with the y-axis indicating net benefit and the x-axis representing threshold probability (g-i)

Discussion

MRI is currently the preferred medical imaging examination for ACL tears. However, the long examination time, numerous contraindications, and high cost limit its clinical application [14]. CT is the preferred emergency examination method for knee joint injuries, mainly used for the diagnosis of fractures and joint dislocations, but has not been clinically applied in the diagnosis of ACL tears. The main reason is that the soft tissue resolution of CT is low, resulting in poor display of pathological manifestations such as bleeding and edema after ACL injury. Moreover, CT images are cross-sectional, and fully displaying the morphology of ACL requires 3D reconstruction, which takes a lot of time. In this study, we first developed a deep learning model based on knee CT, which could achieve accurate segmentation of ACL. It avoided manual ACL reconstruction and segmentation of traditional method. By using radiomics and machine learning techniques, we developed three artificial intelligence models that can accurately diagnose ACL tears. The AUC of SVM, RF and SGD in the independent external validation set were 0.85, 0.86 and 0.81, respectively. This suggested that our model had significant potential clinical application value, especially for patients who could not undergo magnetic resonance imaging due to contraindications and other reasons. In emergency situations of knee injuries, CT is usually a routine examination. Our segmentation and classification model could help diagnose ACL tears based on CT images, which has important clinical significance.

This study use deep learning to construct an ACL segmentation model based on knee joint CT images, and achieved high segmentation accuracy. Previously, researchers [15, 16] had developed various deep learning algorithms for knee CT images segmentation, which could quickly and accurately segment the femur, tibia, patella, and fibula, detect the attachment points of the quadriceps, hamstring, biceps femoris tendons, lateral and medial collateral ligaments, and measure patellar tilt. Cohen et al. [17] developed an artificial intelligence algorithm that could accurately detect knee joint effusion and classify simple effusion and hematoma. In this study, we used 3D nnU-Net [13] with adaptive topology structure as the core framework to construct the ACL segmentation model. 3D nnU-Net preserved the multi-level feature fusion mechanism of U-Net and could dynamically configuring network depth, convolution kernel size, and regularization strategy to achieve parameter self optimization based on data characteristics, avoiding the high dependence of traditional methods on manual parameter tuning and significantly improving the robustness and generalization ability of the model. This framework could fully analyze the three-dimensional spatial correlation of medical images, effectively capture the spatial features of ACL in cross-sectional, coronal, and sagittal planes, and accurately model complex. This ability was not available in traditional single plane segmentation methods or fixed architecture models, which greatly facilitated the accurate ACL morphology analysis, tear assessment, or surgical planning in the next step. Previous studies had confirmed that the 3D nnU -Net architecture exhibited high accuracy in various medical image segmentation tasks. Guo X et al. [18] found that the 3D nnU-Net achieved a DSC of 0.88% and 0.87% for automated segmenting of true and false lumens in CT data of type A aortic dissection, respectively. Yang E et al. [19] used 3D nnU-Net to segment the abdominal CT images of patients with pancreatic cancer. Their results showed that it could effectively segment the pancreatic and accurately measure the pancreatic volume, with DSC reaching 0.81. The results of Wang et al. [20] indicated that the 3D nnU-Net al.so performed well in segmenting uterine fibroids and their surrounding organs, significantly outperforming other models such as HIFUNet, U-Net, R2U Net, ConvUNeXt, and 2D nnU-Net. However, in a patient of the internal testing data, the model incorrectly identified and segmented a small piece of tissue near the hip joint as part of the target tissue. This indicated that after the segmentation was completed, further manual confirmation was necessary in clinical use.

In the construction of radiomics model for ACL tear diagnosis, LASSO helped identify subsets of features with clear pathological associations, such as texture features reflecting the direction of ACL fiber bundles and morphological features characterizing the degree of ACL tearing. The retained 78 features included 18 first-order statistical features, 3 shape features, 44 texture features, and 13 high-order features. These features could provide comprehensive image information such as voxel intensity distribution, overall geometric shape, and internal texture heterogeneity of ACL. The shape features reflected the macroscopic structural information such as the contour and continuity of ACL. When ACL tears, its contour becomes blurred, distorted, wavy, focal thickening or thinning, and even completely interrupted continuity, which could be recognized and quantified on these features. Texture features reflect the grayscale relationships between image pixels or voxels, mainly reflecting microstructural information such as spatial distribution patterns, uniformity, or heterogeneity of signal intensity within the ACL. Acute bleeding, inflammatory edema, and subsequent pathological processes such as fibrous tissue proliferation and mucinous degeneration that occur after ACL tear could significantly alter these texture features. In previous studies [21, 22], it had been proven that radiomics features could effectively reflect macroscopic and microscopic pathological changes in tissues, and were more sensitive than subjective assessments by human eyes. In this study, SVM, RF, and SGD were used for machine learning training, and all three models performed well. These three models have many advantages and have been widely used in medical research and clinical fields. The advantage of SVM [23] lies in its ability to map the original high-dimensional nonlinear image features to a higher dimensional or even infinite dimensional space, effectively capturing tiny information and possessing excellent stability and strong generalization ability. RF [24] has excellent robustness and inherent feature processing mechanisms, which can effectively solve the problem of redundant features and image noise faced by radiomics analysis. It can also automatically identify and sort radiomics features, ensuring the interpretability and accuracy of the model. SGD [25] can update the model in real-time, allowing it to continue evolving. This dynamic update capability can effectively overcome the problem of result drift caused by differences in imaging protocols among different institutions or patient population.

This study had several limitations. Firstly, this study was a retrospective research design and there was a risk of sample selection bias. Secondly, only data from two Hospital s were included, resulting in a relatively small sample size. Finally, the gold standard for ACL tear diagnosis is derived from MRI, not arthroscopic surgical confirmation. This might lead to inherent bias in the labels and the inability to classify specific types of ACL tears. Future work needs a prospective study design, recruiting larger sample sizes of patients from multiple medical centers, and further validating our results using arthroscopic surgery as the final gold standard.

Conclusions

In summary, our research had developed a CT based artificial intelligence system that could quickly perform ACL segmentation and diagnosis of ACL tears. It has high accuracy and convenience, and was of great significance in clinical practice.

Abbreviations

ACL

Anterior cruciate ligament

CT

Computed tomography

MRI

Magnetic resonance imaging

DICOM

Digital imaging and communications in medicine

HU

Hounsfield Unit

GPU

Graphics Processing Unit

SVM

Support vector machine

RF

Random forest

SGD

Stochastic gradient descent

DCA

Decision curve analysis

DSC

Dice Similarity Coefficient

RVE

Relative Volume Error

HD95

95% Hausdorff Distance

ASSD

Average Symmetric Surface Distance

AUC

Area under the curve

ROC

Receiver operating characteristic

PDWI

Proton Density Weighted Imaging

CNN

Convolutional Neural Network

LASSO

Least absolute shrinkage and selection operator

Author contributions

Xiaoli Yu: Conceptualization, Methodology, Formal Analysis, Writing - Original Draft; Qingning Yang: Data Curation, Methodology, Writing-Original Draft; Xingyan Le: Literature search, Investigation; Qingbiao Zhang: Data Collection, Software, Visualization; Yuyin Wang: Formal Analysis, Validation; Junbang Feng: Conceptualization, Supervision, Project administration, Writing - Review & Editing; Chuanming Li: Conceptualization, Funding Acquisition, Resources, Supervision, Writing - Review & Editing.

Funding

This study was funded by (1) Chongqing Key Laboratory of Emergency Medicine Project (2024RCCX05),(2) Natural Science Foundation of Chongqing (CSTB2024NSCQ-KJFZMSX0051), (3) Science and Technology Research Program of Chongqing Education Commission (KJQN202300137).

Data availability

The datasets generated and analysed during the current study are not publicly available due to data protection and privacy regulations but are available from the corresponding author on reasonable request and with appropriate institutional approvals.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki. This retrospective study was granted a waiver of informed consent and has been approved by the Research Ethics Committees of Chongqing University Central Hospital and Chongqing University Fuling Hospital. Ethical Approval Document No. 2025 (104).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.