The study aimed to create a dataset of Iliac CTV scans for the automated detection of MTS using deep learning and radiomics. The dataset used in this study is representative of diverse patient populations, including individuals across different age groups (18–85 years) and genders (52% female, 48% male). While MTS is more prevalent in females, particularly in the 20–50 age group, our gender distribution reflects the inclusion of both primary MTS cases and secondary cases associated with conditions like deep vein thrombosis (DVT), which can affect both genders. Data were collected from multiple clinical settings, such as tertiary care hospitals and community clinics, to ensure variability in clinical presentations and imaging protocols. This diversity enhances the generalizability of our findings and supports the applicability of the proposed methodology to a broad spectrum of patients with MTS. The inclusion and exclusion criteria were designed to ensure that the selected cases were relevant to the research objectives while excluding cases with potential confounding factors or data quality issues. In the Algorithmic Investigation for Automated Detection of May-Thurner syndrome from Iliac CTV imaging via Deep Learning and Radiomics, the criteria for including and excluding participants in the study population were outlined as follows.

Inclusion Criteria:

Exclusion Criteria:

Figure 1A demonstrates an example of a left iliac vein meeting image quality criteria with optimal contrast filling and no artifacts, while Figures 1B,C illustrate excluded cases due to incomplete contrast opacification or imaging artifacts that could compromise algorithmic training. Figure 1D represents a post-iliac vein stenting case, which was also excluded from the study cohort in accordance with our exclusion criteria.

A database of left iliac CTV scans was collected from the CT/MRI department, anonymized to comply with HIPAA regulations, and labeled by experienced radiologists to indicate the presence or absence of MTS. It’s important to note that all patients underwent Digital Subtraction Angiography (DSA) surgery performed by experienced physicians, and DSA was considered the gold standard for confirming the presence of MTS. This dataset comprises a total of 490 cases, with 201 cases showing positive results for MTS and 289 cases with negative results.

The scanning protocol for Iliac Vein CT Venography (CTV) involves several important steps. Firstly, patient preparation is crucial, including ensuring the patient is well-hydrated, obtaining informed consent, and checking for medical history and contrast agent allergies. Following the preparatory procedures, the patient is positioned supine on the table of the CT scanner. An intravenous administration of a contrast agent is then performed using a power injector, delivering the medium at a rate of 3 to 4 mL per second, generally utilizing a non-ionic iodinated contrast substance. The choice of scanner is typically a Multidetector CT (MDCT) for CTV. Specific scanner parameters are set to achieve high-quality images. The parameters established for the imaging protocol consist of a slice thickness that ranges from 1 to 2 mm. The imaging technique captures the entire length of the iliac vein, from the common iliac vein through to the femoral vein. In general practice, the tube voltage is fixed at 120 kV, whereas the tube current is adjusted based on the individual’s physical attributes, typically falling between 200 and 300 mA. To ensure effective data collection, a pitch between 0.75 and 1.5 is employed, and a reconstruction interval of 1 mm is utilized to achieve images of superior quality. The scan delay is determined to ensure optimal timing for image acquisition, usually occurring around 20–30 s after contrast injection. The subsequent steps in post-processing encompass multi-planar reformatting (MPR) and maximum intensity projection (MIP), which facilitate the visualization of the iliac veins from various angles, and both arterial and venous phase images are acquired for comprehensive visualization. Finally, the acquired images are carefully evaluated for any venous pathology or abnormalities, such as deep vein thrombosis (DVT) or compression syndromes. As shown in Figure 2, we present the CTV images from our research dataset along with their manifestations in DSA. We separately display the appearances of MTS-positive CTV and DSA.

The dataset underwent several preprocessing steps to ensure consistency and enhance model performance. First, all Computed Tomography Venography (CTV) images were resampled to a uniform voxel size of 0.6 mm³ to standardize spatial resolution across scans. Image normalization was performed by scaling the intensity values to a range of [0, 1] using min-max normalization, which reduces variability in pixel intensity due to differences in imaging protocols. To address class imbalance in the dataset, we applied data augmentation techniques, including random rotations (±10°), horizontal and vertical flips, and slight translations (±5% of image dimensions). These augmentations were applied only to the training set to increase its diversity and improve model generalization. Additionally, regions of interest (ROIs) containing the iliac veins and surrounding structures were manually annotated by experienced radiologists to provide ground truth labels for model training and evaluation.

The 10-fold cross-validation method is employed to both train and assess the performance of our deep learning-driven automatic vein segmentation algorithm. After 10-fold cross-validation, each case obtains a segmentation result from a deep learning model. We juxtapose these findings against the manual segmentation outcomes provided by seasoned medical professionals. Furthermore, we compute various segmentation performance evaluation metrics to gauge the efficacy of the model’s segmentation capabilities. Subsequently, this segmentation data is employed to construct a radiomics model aimed at identifying the presence of MTS syndrome.

The initial phase of the conversion procedure involved the conversion of Digital Imaging and Communications in Medicine (DICOM) files, sourced from a PACS batch, into the NIFTI format. Under a rigorous multi-stage annotation protocol, all left iliac vein structures were meticulously delineated using ITK-SNAP software (version 4.0.0). This critical annotation process was conducted through collaborative verification by three senior attending physicians (with over 10 years of vascular imaging experience) and one biomedical engineer, with final annotations requiring unanimous approval from all three physicians to ensure maximal inter-observer consistency. Technical specifications included: (1) Optimized visualization parameters set at a window width of 350 Hounsfield units (HU) and window level of 60 HU; (2) Systematic removal of osseous components through multi-planar reconstruction analysis to establish gold-standard references for segmentation model training. This consensus-driven annotation framework significantly reduced intra- and inter-operator variability (achieving Dice similarity coefficients > 0.95 in validation tests), thereby enhancing the anatomical accuracy and clinical reliability of the subsequent automated segmentation models.

Image segmentation serves as a fundamental aspect of computer vision, entailing the division of images into significant segments that align with different objects or areas of interest. The UPerNet architecture tackles the intricacies associated with semantic segmentation by seamlessly combining multi-scale features and leveraging contextual information to enhance the accuracy of segmentation (34).

The detailed depiction of the segmentation model is illustrated in Figure 3. The UPerNet framework consists of three primary elements: (1) a Feature Pyramid Network (FPN) designed for the extraction of features across multiple scales, (2) a Global Aggregation Module (GAM) designed to improve contextual understanding, and (3) a Perceptual Enhancement Module (PEM) focused on the enhancement of features.

The FPN module operates through a dual-pathway system that integrates both bottom-up and top-down processes, enabling the network to capture features across different scales. This dual mechanism allows for the merging of detailed nuances with broader contextual elements, making the FPN essential for effectively managing objects of diverse sizes while maintaining spatial coherence.

The Global Attention Module (GAM) enhances UPerNet’s capacity to extract extensive contextual dependencies by integrating information across multiple scales. By employing global context pooling and convolutional operations, the network is equipped to identify complex relationships among various parts of an image. This component serves as a critical tool for achieving precise segmentation within complicated scenes, thereby playing a pivotal role in the overall effectiveness of UPerNet.

The Perceptual Enhancement Module (PEM) acts as a pivotal platform for the enhancement of features and the enrichment of semantic content. Drawing inspiration from the mechanisms underlying human visual perception, the PEM incorporates perceptual loss functions that effectively steer the training pathway of the network. By ensuring that the features obtained are in harmony with human perceptual evaluations, the module significantly enhances UPerNet’s ability to generate segmentation results that are semantically coherent and meaningful.

The application of Focal Loss in the training of segmentation models presents an effective approach to tackle the issue of class imbalance while simultaneously improving boundary precision. By adaptively modifying the contributions of loss according to the confidence of predictions, Focal Loss directs the network’s attention toward more difficult areas, resulting in enhanced segmentation outcomes. The positive outcomes observed across diverse datasets underscore the potential of Focal Loss to elevate the state-of-the-art in image segmentation (35).

The study utilized an NVIDIA RTX A5000 graphics processing unit, which possesses a memory capacity of 24 GB. For the software environment, Python version 3.6 was implemented, supplemented by libraries including Pytorch 0.4.1, OpenCV, Numpy, and SimpleITK. The input data comprised thin-layer computed tomography (CT) images and the corresponding labels (36). The outcome achieved was a segmented representation of the iliac vein in CTV. In the initial stages of training, a batch size comprising 16 images was employed for every iteration, and the learning rate was adjusted to 10⁻⁴, with a cumulative total of 200 training epochs.

The objective assessment approach utilized the Dice coefficient from the test dataset to evaluate the performance of the segmentation model.

To assess the segmentation performance through subjective evaluation, a 10-fold cross-validation method was employed, and the model that had been trained was subsequently utilized to forecast the data pertaining to the remaining cases.

A stratified random sampling method was employed to distinguish between the MTS and non-MTS cohorts. The cases were allocated proportionally into a training dataset and an external validation dataset at a ratio of 7:3. A total of 1,794 distinct features were extracted utilizing the PyRadiomics framework, which can be broadly classified into several categories: first-order features, shape descriptors, and various texture matrices, including the Gray Level Co-occurrence Matrix (GLCM), Gray Level Size Zone Matrix (GLSZM), Gray Level Run Length Matrix (GLRLM), Neighboring Gray Tone Difference Matrix (NGTD), and Gray Level Dependence Matrix (GLDM). The filtering parameters applied encompassed wavelet transformations, Laplacian of Gaussian (LoG), Square, Square Root, logarithmic, exponential, gradient, along with Local Binary Pattern in two dimensions (LBP2D) and three dimensions (LBP3D). The bin width was established at 25, and the resampled voxel dimensions were set to 3 × 3 × 3. For the LoG filter, the kernel sizes ranged from 1 to 5.

The characteristic derived from radiomic analysis dataset was subjected to a preprocessing phase, during which outliers and missing values were substituted with the median. This was subsequently accompanied by standardization of the data to mitigate the effects of dimensionality. Furthermore, we employed the minimum-redundancy maximum-relevance (mRMR) algorithm to pinpoint the nine most crucial features that exhibit a robust correlation with MTS. In this study, we retained 10 features using the mRMR algorithm. The Least Absolute Shrinkage and Selection Operator (LASSO) regression methodology was employed to identify the non-zero coefficient features that exhibit the highest relevance to the diagnosis of MTS from the features retained. Subsequently, the radiomics score (Rad-score) for each MTS was computed by executing a linear combination of the weighted coefficients associated with the selected features. The procedural framework for radiomics analysis and the development of the segmentation model is illustrated in Figure 3.

All analyses performed in this study were carried out using R software (http://www.Rproject.org, version: 3.6.1). To compare continuous and categorical variables, the t-test and Chi-square test were utilized, respectively, with a p value threshold of less than 0.05 considered statistically significant. The diagnostic efficacy of the radiomic signature was evaluated through receiver operating characteristic (ROC) curve analysis. Furthermore, the Hosmer-Lemeshow test was employed to assess the goodness of fit of the radiomic signature.

In light of the ratio between the MTS group and the non-MTS group, we employed a 10-fold cross-validation approach to both train and assess our deep learning-driven automatic vein segmentation algorithm. Figure 4a depicts the loss trajectory associated with the training phase of the model, whereas Figure 4b displays the fluctuations in Dice coefficients observed within the testing cohort. The UPerNet models that underwent training were subsequently utilized to segment all 490 cases for the purpose of external validation. These masks were subsequently employed in the creation and assessment of the radiomic signature. The results were compared against labels that were manually created. The UPerNet model achieved an average Dice coefficient of 0.925 (95% CI: 0.875–0.961) after completing the 10-fold cross-validation procedure.

The nine most predictive features were selected using mRMR and LASSO algorithms, Subsequent to the creation of a radiomic signature via logistic regression employing the selected characteristics, the corresponding Rad-score formula can be referenced in Figure 5. The radiomic signature demonstrated notable diagnostic performance for MTS, attaining an area under the curve (AUC) of 0.891 (with a range of 0.825 to 0.956) within the training cohort. In the validation cohort, it achieved an AUC of 0.892 (spanning from 0.793 to 0.991), as depicted in Figure 6. The detailed findings of the receiver operating characteristic (ROC) analysis are presented in Table 1. The results from the Hosmer-Lemeshow test revealed that the radiomic signature showed no evidence of overfitting in the training cohort, the internal validation cohort, and the external validation cohort, with all p-values exceeding 0.05. To clarify the importance of specific features and their combined impact on diagnostic performance, SHAP summary plots illustrating the radiomic signature were generated, as shown in Figure 7.

The original_shape_MinorAxisLength (coefficient: −0.8) provided direct morphometric evidence of left iliac vein compression, with its negative weight reflecting the diagnostic significance of lumen diameter reduction—a finding consistent with catheter venography standards (39). Complementing this, log-sigma-4-0-mm-3D_firstorder_90Percentile (coefficient: +0.7) quantified perivascular fibrosis through high-intensity signals at a 4 mm spatial scale, correlating with histopathological collagen deposition (40). Notably, the exponential_glcm_1mc2 (coefficient: −1.1) further resolved microstructural heterogeneity in fibrotic venous walls by measuring local texture contrast. Elevated values of this feature aligned with asymmetric collagen distribution patterns observed in Masson’s trichrome-stained specimens (40).

The exponential_glcm_Autocorrelation (coefficient: +0.6) captured macroscale flow turbulence through its quantification of pixel intensity dependencies. Its strong negative weight indicated disrupted flow regularity, a hallmark of extrinsic iliac artery compression. At a finer scale, log-sigma-5-0-mm-3D_gldm_DependenceVariance (coefficient: −0.4) mapped flow stagnation zones by analyzing gray-level dependencies, with increased variance values matching phase-contrast MRI evidence of retrograde flow (p < 0.01) (41). The logarithm_glcm_ldn (coefficient: −0.9) enhanced early detection of hemodynamic shifts by amplifying subtle intensity variations in pre-stenotic regions, achieving 89% sensitivity for incipient intimal hyperplasia in our cohort—a critical advancement for early-stage MTS diagnosis (42).

Thrombus-specific features demonstrated exceptional discriminatory power. The wavelet-HHH_dlszm_LargeAreaHighGrayLevelEmphasis (coefficient: −0.1) identified hyperdense thrombus cores through 3D wavelet decomposition, with a 92% concordance rate against contrast-enhanced ultrasound (43). Concurrently, gradient_firstorder_Kurtosis (coefficient: −0.15) detected acute thrombus margins via sharp intensity transitions, showing 40% higher sensitivity than manual ROI analysis (p = 0.003) (44). The exponential_dlszm_ZonePercentage (coefficient: +1.2) further stratified thrombus maturity, where reduced homogeneity (negative weight) correlated with histopathological evidence of lytic reorganization (AUC = 0.88) (45). Radiomic features, encapsulating the essence of medical image analysis, were judiciously extracted, thereby setting the stage for the development of a robust radiomic signature tailored specifically for MTS diagnosis (46). The radiomic signature, which underscores its diagnostic effectiveness, consistently produced elevated AUC values across both the training and validation groups (47, 48).

In summation, the judicious application of deep learning and radiomics modalities to Iliac CTV scans portends considerable promise for elevating the diagnostic paradigm for MTS. These sophisticated techniques stand poised to augment diagnostic efficiency and accuracy in the identification of this intricate vascular syndrome, thereby effecting tangible advancements in the realm of patient care.