Quantitative assessment and risk stratification of random acute pulmonary embolism cases using a deep learning model based on computed tomography pulmonary angiography images
Original Article

Quantitative assessment and risk stratification of random acute pulmonary embolism cases using a deep learning model based on computed tomography pulmonary angiography images

Yang Qiao1,2 ORCID logo, Yaozong Gao3, Yanbo Chen3, Xiaodan Ye1,2, Cheng Yan1,2, Mengsu Zeng1,2

1Department of Radiology, Zhongshan Hospital Affiliated to Fudan University, Shanghai, China; 2Shanghai Institute of Medical Imaging, Shanghai, China; 3United Imaging Intelligence, Shanghai, China

Contributions: (I) Conception and design: X Ye, M Zeng; (II) Administrative support: M Zeng; (III) Provision of study materials or patients: Y Qiao, Y Gao, Y Chen, C Yan; (IV) Collection and assembly of data: Y Qiao, Y Gao, Y Chen, C Yan; (V) Data analysis and interpretation: Y Qiao, Y Gao, Y Chen, X Ye, M Zeng; (VI) Manuscript writing: All authors; (VII) Final approval of manuscript: All authors.

Correspondence to: Mengsu Zeng, PhD. Department of Radiology, Zhongshan Hospital Affiliated to Fudan University, No. 180 Fenglin Road, Xuhui District, Shanghai 200032, China; Shanghai Institute of Medical Imaging, Shanghai, China. Email: zengmengsu20210116@163.com.

Background: Computed tomography pulmonary angiography (CTPA) is the gold standard for the diagnosis of pulmonary embolism (PE). The semi-quantitative clot burden scoring based on imaging is related to the risk stratification and prognosis of acute PE, but it cannot be widely applied in the clinic due to the difficulty of calculation. This study developed a high-quality VB-Net deep learning (DL) model combined with Transformer, which can detect PE from images and automatically calculate the clot burden score (CBS). The aim of this study was to help patients via earlier diagnosis, risk stratification, and determination of treatment plans, thereby improving prognosis, as well as alleviate the burden on radiologists. To our knowledge, no related studies have been reported.

Methods: A retrospective inclusion of 2,424 CTPA examination cases (44% male) were conducted to train and test the VB-Net DL model for the detection of PE and to evaluate the clot burden volume and scoring. Area under the curve (AUC), and sensitivity and specificity on the case or clot level were used to evaluate the model’s performance. Random CTPA data from Zhongshan Hospital Affiliated to Fudan University (30 cases with acute PE, 40 cases without PE) were applied to test the relationship between the clot burden automatically calculated by the model and the Qanadli score determined manually, as well as other imaging parameters.

Results: The performance of the VB-Net DL model on the testing set had an AUC of 0.972 based on the case level. The sensitivity at the operational point of the model threshold selected was 94.6% [95% confidence interval (CI): 0.8650–0.9828], and the specificity was 89.4% (95% CI: 0.8407–0.9308). In the random CTPA examinations from this research center, the model’s sensitivity based on the case was 76.67% (95% CI: 0.5880–0.8848), the specificity was 95.00% (95% CI: 0.8261–0.9950), the positive predictive value (PPV) was 92.00%, and the accuracy was 87.14%. On the clot-based level, the sensitivity was 84.43%, the PPV was 87.29%, and the false positive rate was 0.19 per case. The clot burden volume and score automatically measured by the model were significantly correlated with the manually determined Qanadli score (r=0.866, P<0.001 and r=0.899, P<0.001, respectively). The severity grading of the CBS groups was consistent with the degree of right ventricular dilation.

Conclusions: The VB-Net DL model based on CTPA could conveniently and efficiently detect and quantitatively evaluate PE.

Keywords: Acute pulmonary embolism (APE); computed tomography pulmonary angiography (CTPA); deep learning model (DL model); automatic clot burden score (automatic CBS)


Submitted Jul 11, 2024. Accepted for publication Feb 04, 2025. Published online Feb 26, 2025.

doi: 10.21037/qims-24-1412


Introduction

Pulmonary embolism (PE) is one of the three most common cardiovascular diseases, characterized by high incidence, high mortality, and high misdiagnosis rates, with over 100,000 deaths worldwide each year (1,2). Most deaths caused by PE occur within the first few hours and early initiation of treatment is associated with better prognosis (3), making early diagnosis particularly crucial. Currently, computed tomography pulmonary angiography (CTPA) is the gold standard for PE diagnosis, with extremely high sensitivity and specificity, and is also the primary imaging tool for patient risk and severity assessment (1). In recent years, with the widespread use of CTPA, the increased rate of false positives has become a new issue, adding to the medical burden. Clot burden is related to the risk stratification and prognosis of acute pulmonary embolism (APE) (4-6). The Qanadli score (7) is currently the most cited semi-quantitative method for evaluating clot burden in CTPA, capable of reflecting the proportion of obstruction in the pulmonary vascular tree. However, due to the time-consuming calculation and reliance on the clinical experience of radiologists, this method has not been widely applied in clinical practice. Deep learning (DL) models based on convolutional neural networks (CNN) are increasingly being applied to the automatic identification and assessment of PE. Fink et al. (8) applied a DL model to identify PE and automatically calculate the clot burden score (CBS), but their study was based on structured reports rather than images; Liu et al. (9) applied a U-net model to automatically measure the clot volume in CTPA images; however, to date, no studies have applied DL models to automatically measure the CBS of PE based on CTPA images. Right ventricular function can also be assessed on CTPA images by measuring the ratio of the right ventricular diameter (RVd) to the left ventricular diameter (LVd). This ratio is an independent risk factor for adverse outcomes in patients with APE (10) and can serve as a prognostic indicator.

In recent years, DL methods in the field of medical image segmentation are most notably represented by U-Net and V-Net. The V-Net model was initially used to segment the prostate from magnetic resonance imaging (MRI) images using an end-to-end fully convolutional network. It combines the characteristics of U-Net and residual networks, offering smoother gradient flow, which is easier to optimize and converge, surpassing many classic CNNs in terms of accuracy in image segmentation. An improved version based on V-Net, VB-Net, has been validated on some internationally common datasets and the datasets from the International Symposium on Biomedical Imaging (ISBI) Segmentation of Thoracic Organs at Risk in Computed Tomography images (SegTHOR) 2019 challenge, demonstrating high accuracy, speed, and robustness (11). Compared to V-Net, VB-Net achieves faster organ segmentation and consumes significantly less GPU memory. At the same time, the VB-Net model, by performing spatial convolution on the dimensionality-reduced feature images, can significantly reduce the number of model parameters. Transformers can simultaneously consider the apparent features and the positional information of the image patches, offering a more global perspective. Through the self-attention mechanism, they can comprehensively consider the relationships between different pulmonary artery (PA) image patches, ensuring the continuity of the PA. At the same time, through positional encoding, they are more conducive to the grading of the PA.

Therefore, this study applied VB-Net combined with Transformer to develop a model for automatically detecting PE in CTPA images and calculating (semi-)quantitative CBSs. To our knowledge, no related studies have been reported yet.

The aim of this study was to develop a high-quality model suitable for clinical application, which can help patients with earlier diagnosis, risk stratification, and determination of treatment plans, thereby improving prognosis. Another objective of the model was to alleviate the burden on radiologists. We present this article in accordance with the TRIPOD + AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-24-1412/rc).


Methods

DL model

Data

A total of 2,424 CTPA examinations were collected, mainly from the public dataset Kaggle, and some from Zhongshan Hospital Affiliated to Fudan University (from 2018 to 2022). The collected data exhibited a relatively high image quality and a balanced gender distribution, with 44% male and 56% female participants. The age distribution ranged from 18 to 100 years. Computed tomography (CT) equipment manufacturers included Siemens (Erlangen, Germany), TOSHIBA (Tokyo, Japan), Philips (Amsterdam, Netherlands), GE (Chicago, IL, USA) and United Imaging Healthcare (UIH; Shanghai, China), and these suppliers held the majority of the CT equipment market. The images primarily consisted of a layer thickness of 1 mm, while also including 0.625 and 1.25 mm.

These 2,424 cases were randomly divided into a training set, a tuning set, and a testing set, with an approximate ratio of 8:1:1. This included 1,900 cases in the training set, 262 cases in the tuning set, and 262 cases in the testing set.

Annotation

The annotation work was composed of trained annotators (two individuals), reviewing personnel (two mid-level experienced doctors), and an arbitrator (one senior experienced doctor). After relevant training, the two professional annotators annotated a batch of images, and the reviewing personnel evaluated the theoretical knowledge and annotation accuracy to select the best annotator to be the annotation team leader. Firstly, annotators labeled the data, which was followed by inspection and modification by the annotation team leader. If there were doubts about the results from the annotation team leader, it could be reported to the reviewing personnel for judgment. If there was a disagreement between the two reviewing personnel, it was then handed over to the arbitrator to finally determine the modification opinion, and then it was returned to the annotation team leader for modification. All annotations followed the same protocol: excluding artifacts, any filling defects or interruptions in the PA were considered clots. All PA and clots greater than 1 mm needed to be annotated. In some cases, the pulmonary veins were also annotated according to the same rules (Figure 1).

Figure 1 Manual annotation for pulmonary arteries, pulmonary veins, and clots. (A) Original CTPA image; (B) CTPA image with annotation; (C) annotated VR image. Red labels represent pulmonary arteries, green labels represent pulmonary veins, blue labels represent clots. CTPA, computed tomography pulmonary angiography; VR, volume rendering.

The specific steps of annotation were as follows:

  • Open-source software platforms [3D Slicer (https://www.slicer.org/) or ITK-SNAP (http://www.itksnap.org/)] were used;
  • The window width and window level were adjusted to the soft tissue window, commonly set as a window width of 400 and window level of 40, with slight adjustments based on the varying degree of vascular enhancement in different images;
  • The smearing or outlining tool was used to delineate the PA and clots areas, assigning different labels to annotate the PA and clots separately;
  • The annotated mask was saved.

Automatic data collection

The overall workflow of VB-Net is shown in Figure 2, demonstrating its capability to automatically segment CTPA images, label clots, measure clot volume, and classify the degree of PA occlusion and clot location.

Figure 2 Overall workflow of VB-Net.

VB-Net consisted of two paths, serving as an improved three-dimensional CNN that combined V-Net with a bottleneck structure to reduce parameters and accelerate network convergence (11,12). The first path was a contraction path, including down-sampling and convolution operations to extract global image features. The second path was an expansion path, including up-sampling and convolution operations to integrate fine-grained image features. Our VB-Net replaced the 5×5×5 convolution operation in V-Net with a bottleneck structure. The first convolution layer used a 1×1×1 convolution kernel, reducing the number of channels in the feature map. The second convolution layer performed spatial convolution with a 3×3×3 kernel. The final convolution layer used a 1×1×1 convolution kernel to increase the printed feature channels back to their original size.

Segmentation of PA and clots

We adopted a two-stage segmentation strategy. The first stage network was trained using the training set, taking the raw images as input and producing a rough segmentation of the PA and veins regions, encompassing the entire vascular network. This stage aimed to learn the overall topological structure of the pulmonary vessels, making it easier to distinguish between pulmonary arteries and veins in subsequent stages and ensuring the continuity of the PA. On this basis, the centerline of the PA was extracted to obtain the PA skeleton.

The output of the first stage was processed and combined with the original image to serve as input for the second-stage network. The second stage involved extracting masks for the PA and clots. A classifier was then applied to remove false-positive clots, thereby improving accuracy. The remaining clots were merged with the PA mask to produce the final output, which included two labels: PA and clot. This two-stage network strategy not only focused on the target regions but also improved the detection performance of clots and the reconstruction quality of the pulmonary arteries (Figure 3). The results of the ablation study are shown in Table S1.

Figure 3 Flowchart of the two-stage segmentation algorithm.

Classification and segmentation of clot obstruction degree

In the process of calculating the Pulmonary Artery Obstruction Index (PAOI), it was necessary to assess the degree of obstruction of the PA (clots causing complete obstruction of the PA or clots causing incomplete obstruction of the PA), assigning different scores. We defined a clot as a “completely obstructive clot” when it occupied the entire cross-sectional area of PA branches; otherwise, it was defined as a “partially obstructive clot”. We used a DL classification network for this purpose. The input to the classification network was the segmented PA and each clot mask from the previous step. The output was a binary classification indicating whether the clot completely or partially occludes the PA. For clots that completely occlude the PA, a DL segmentation network further divides them into non-occluded (a) and occluded parts (b) (Figure 4).

Figure 4 Flowchart of the classification and segmentation of clot obstruction degree.

PA vessel grading

The PA and clot masks obtained in 1.3.1 were combined to create a binary mask, denoted as M. This mask, along with the original image, was then processed through a segmentation network to obtain PA grading results. The grading results included five levels: PA trunk level (1 label), left and right main PA level (2 labels), lobar level (5 lobe labels), segmental level (20 segment labels), and subsegmental level. This study introduced a segmentation model that integrated the apparent features and positional information of the PA. Firstly, the centerline skeleton of the PA was extracted from the binary mask M. Fixed-size blocks were selected along this centerline, ensuring that the center point of each block lay on the PA centerline. This sampling strategy ensured a high area proportion of the PA in each block. Simultaneously, there was slight overlap between adjacent blocks to ensure no omission of any PA region. Subsequently, a CNN served as the encoder to extract the apparent features of the blocks. For a sampled block with a world coordinate of pwi,i{x,y,z}, the position encoding was defined as: t=(pwicmini)/(cmaxicmini). Here, cmini and cmaxi represented the minimum voxel coordinate point (0,0,0) and the maximum (h,w,d) of the minimum bounding box (size h×w×d) containing the binary mask M, respectively, corresponding to world coordinates.

The CNN encoder processed the block p and extracted a one-dimensional vector of apparent features: e=Encoder(p). One advantage of using a CNN to extract apparent features is the reduction in the dimensionality of feature information. This not only retains expressive feature capabilities but also reduces the number of network parameters, improving the computational efficiency of the DL model. Storing blocks in sequence, inspired by the Transformer model used in natural language processing for machine translation tasks, facilitates the representation of spatial relationships between blocks (13).

The Transformer consists of multiple identical layers, each comprising an encoder module and a decoder module. Both the encoder and decoder consist of a self-attention layer and a feedforward neural network. Additionally, the decoder includes an extra encoder-decoder attention layer between the self-attention layer and the feedforward layer. The self-attention layer lacks the ability to capture positional information of input blocks. To address this, the position encoding t is embedded into the vector. In this study, after passing through a multilayer perceptron (MLP), t was added to the apparent feature vector: ep=eMLP(t). Here, ⊕ denoted matrix addition, ep was the apparent feature vector including positional information, and MLP stood for multilayer perceptron. Using a stacked Transformer with L layers, the network input-output eL was established: eL=Transformer(ep). Finally, the decoder transformed evL into pixel-level classification results, corresponding to PA grading masks.

(Semi-)quantitative evaluation of PA embolism

The PAOI was based on the Qanadli scoring system (7). The CBS calculation formula was as follows: CBS=(n×d)÷40×100%, where n represents the number of segment-level pulmonary arteries included in the position of the clot, d represents the degree of PA obstruction.

Specifically, the maximum value of n was 20, corresponding to the 20 segment-level pulmonary arteries. A score of 1 was assigned when a clot appeared in a segmental artery. When clots appeared in arteries at the lobar level or above, the score was equal to the number of PA segments involved (for example, a score of 5 was assigned if a clot appears in the right lower lobe artery, which contains 5 PA segments). The total score could not exceed the maximum value of the region (for example, the left main PA contains 10 PA segments, so the maximum score is 10). d was taken as a binary variable. When a clot completely occluded the vessel, d was assigned a value of 2, and when the clot was not completely occluded, d was assigned a value of 1. Subsegmental clots were considered a partially occluded segmental artery and were assigned a score of 1. In summary, the maximum score of (n×d) is 40 (Figure 5).

Figure 5 Pulmonary artery vessel grading and calculation of pulmonary artery obstruction index. n represents the number of segment-level pulmonary arteries included in the position of the clot, d represents the degree of PA obstruction (1 for partial, half-filled circle or 2 for completely, filled circle). Red labels represent pulmonary artery trunk, green labels represent Left main PA, blue labels represent Right main PA, yellow labels represent the degree of PA obstruction. MPA-R/L, right and left main PA; ML, right middle lobe; LLL, left lower lobe; LUL, left upper lobe; RLL, right lower lobe; PA, pulmonary artery; RUL, right upper lobe; S, segmental.

The formula for calculating clot burden volume (CBV) was: CBV=(Vi). Here, V represented the volume of a clot. The CBV was automatically measured by the DL model, with the unit being cubic millimeters (mm³).

Other relevant data collection

The VB model automatically reconstructs the four-chamber heart view, which passes through the apex of the heart and is perpendicular to the interventricular septum. The model automatically measures RVd and LVd, defined as the maximum vertical distance from the free wall of the ventricle to the interventricular septum. RVd/LVd ratio ≥1 was defined as right ventricular dilation (RVD), as shown in Figure 6.

Figure 6 Schematic diagram of automated measurement of left and right ventricular diameters on reconstructed four-chamber heart level. The diameter of the right ventricle is 43.0 mm, the diameter of the left ventricle is 49.5 mm, the RVd/LVd ratio is 0.869. LVd, left ventricular diameter; RVd, right ventricular diameter.

Application of the model in random CTPA examinations

The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This study was approved by the Institutional Review Board of Zhongshan Hospital Affiliated to Fudan University (No. B2021-555). The requirement for informed consent was waived due to the retrospective nature of the study. All patient data used in this study were de-identified, ensuring that no personally identifiable information was included. We randomly selected cases in March 2023 at Zhongshan Hospital Affiliated to Fudan University, where patients were clinically suspected of having a PE and underwent CTPA. The cases were diagnosed by two radiologists with over eight years of experience. A total of 30 cases were selected with APE and 40 cases without PE. The exclusion criteria were as follows: (I) age less than 18 years; (II) the DL model could not perform the calculation; (III) history of anticoagulant treatment within the past three months. Ultimately, 32 males and 38 females were included, with an age distribution ranging from 42 to 85 years (Figure 7).

Figure 7 A flowchart detailing of selecting participants. APE, acute pulmonary embolism; CTPA, computed tomography pulmonary angiography; PA, pulmonary artery; PE, pulmonary embolism.

CT examination

The CTPA examination utilized contrast-enhanced CT scanning (United Imaging 960+). A non-ionic iodinated contrast agent (containing 370 g/L of iodine) was injected via the right cubital vein using a Medrad high-pressure injector (Bayer HealthCare, Leverkusen, Germany) at an injection rate of 3–4 mL/s, with a total volume of 20–40 mL (adjusted according to the patient’s body weight). After the contrast agent injection, 15 mL of 0.9% sodium chloride solution was flushed in at the same flow rate. The patient was positioned supine with feet first, and bolus-tracking automatic triggering scan technology was used, with a region of interest (ROI) set at the main PA trunk, and the triggering threshold was set at 120 HU. During a single breath-hold, a scan of the PA from the lung apex to the bilateral diaphragmatic margins was performed. The scanning parameters were as follows: gantry rotation time of 0.25 s per rotation, a detector of 0.5×160 rows, tube voltage of 100–120 kV, automatic tube current of 100 mAs, field of view of 250mm × 250 mm, and an acquisition matrix of 512×512. The CT data obtained were processed on a workstation, uploaded in Digital Imaging and Communications in Medicine (DICOM) format, and three-dimensional volume rendering (VR) images were obtained using surface rendering technology. Subsequently, various vascular images, multiplanar reconstruction (MPR) images, and maximum intensity projection (MIP) images were acquired.

Manual diagnosis and scoring

The diagnosis of APE followed the guidelines of the European Society of Cardiology (ESC) and the European Respiratory Society (ERS) from 2019 (1). Excluding exercise or flow artifacts, uneven filling of the contrast agent, the presence of an intraluminal filling defect, or a sharp cutoff in the PA can be diagnosed as PE (14). The number of clots, their location, and semi-quantitative scoring were determined by two physicians of the department of radiology with over eight years of experience, reading the images separately. If there was disagreement or uncertainty between the readers, a physician with more than 15 years of experience in thoracic imaging would make the final decision, which would be recorded as the ultimate result. The scoring was based on the Qanadli scoring system (7), defining continuous filling defects as one clot, recording the grade of the PA corresponding to the proximal end of the clot and the total number, defining the cutoff of the PA opacification as ‘complete occlusion’, and the rest as ‘partial occlusion’ (Figure 5). At the same time, the image characteristics were rated as first, second, or third grade, representing excellent, good, or poor image quality, respectively. The readers were not informed of the patients’ current conditions and prognosis, and they were allowed to adjust the window settings to maximize the contrast between the clot and the PA blood. A comprehensive assessment was made based on axial images, combined with coronal and sagittal vascular reconstructed images.

Statistical analysis

Statistical analysis was conducted using the software SPSS 20.0 (IBM Corp., Armonk, NY, USA). All data tests were two-tailed, and a P value of less than 0.05 was considered statistically significant. Continuous variables were presented as mean ± standard deviation. The model’s sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, false positive rate (number of incorrectly identified positive clots/total number of cases) and the 95% confidence interval (CI) of sensitivity and specificity were calculated. Area under the curve (AUC) was calculated for VB-Net DL model to predict the incidence of PE, as was optimal cutoff value with a maximum sensitivity and specificity model. The correlation between automatic CBS, CBV, and manual Qanadli scoring was assessed using Spearman’s test (r). Bland-Altman plots were used to assess consistency between the scoring data. The Mann-Whitney test was used to compare the differences in RVd/LVd ratio between scoring groups.


Results

Model performance on testing set

The model’s AUC for identifying PE cases based on CTPA was 0.972 (Figure 8). After comprehensive consideration, the operating point was selected with a sensitivity of 94.6% (95% CI: 0.8650–0.9828) and a specificity of 89.4% (95% CI: 0.8407–0.9308). On the clot-based level, the final selected operating point had a sensitivity of 79.3% (95% CI: 0.7462–0.8337), a PPV of 82.7%, and an average of 0.23 false positive clots per case.

Figure 8 The receiver operating characteristic curve of VB-Net model for identifying pulmonary embolism on patient level. AUC, area under the curve; ROC, receiver operating characteristic.

Model application in CTPA examination

Out of 70 cases, 30 were positive for APE, including 14 males (46.67%) and 16 females (53.33%), with an average age of 63.7±14.67 years, there were a total of 122 clots. Among them, 11 cases had image quality rated as first grade, 12 cases as second grade, and 7 cases as third grade. The 40 negative cases for PE included 18 males (45.00%) and 22 females (55.00%), with an average age of 64.1±7.99 years. Among them, 16 cases had image quality rated as first grade, 16 cases as second grade, and 8 cases as third grade. The DL model detected 25 cases of PE, with 23 true and 2 false positives cases. It detected 118 clots, with 103 true and 15 false clots. Thus, the sensitivity based on cases was 76.67% (95% CI: 0.5880–0.8848), specificity 95.00% (95% CI: 0.8261–0.9950), PPV 92.00%, NPV 84.44%, and accuracy 87.14%. The sensitivity based on clots was 84.43%, PPV 87.29%, and the false positive rate was 0.21 per case. The detection based on the location of the clot is shown in Table 1 and Figure 9. There were seven PE cases missed by the model, six of which had a single subsegmental PA clot, and one case had two subsegmental PA clots. The most common causes of misdiagnosed negative cases were superior vena cava contrast artifacts, lung cancer, and compression of the PA by enlarged lymph nodes.

Table 1

Location of the most proximal clot in pulmonary arteries on the deep learning model

Artery location CT scans Clots
TP (n=23) FN (n=7) Total (n=30) Sensitivity (%) TP (n=103) FN (n=19) Total (n=122) Sensitivity (%)
Main 11 0 11 100 16 0 16 100
Lobar 2 0 2 100 13 0 13 100
Segmental 8 0 8 100 37 2 39 94.87
Subsegmental 2 7 9 22.22 37 17 54 68.52

CT, computed tomography; TP, true positive; FN, false negative.

Figure 9 The detection of deep learning model based on the location of the clot. (A) Total and true positive numbers of CT scans. (B) Total and true positive numbers of clots. CT, computed tomography.

The correlation between CBS and Qanadli scores was r=0.899, P<0.001. The correlation between CBS and CBV was r=0.960, P<0.001. The correlation between QD and volume was r=0.866, P<0.001. The mean difference between the model CBS score and the manual Qanadli score was 0.92 (95% CI: −14.76 to 16.60), with a standard deviation of 8.00, where the unit is (%). The consistency and correlation between the two scores were both high (Figure 10).

Figure 10 The consistency and correlation assessment. (A) Bland-Altman plot between CBS and QD. (B) The correlation between CBS and QD, r=0.899, P<0.001. (C) The correlation between QD and CBV, r=0.866, P<0.001. (D) The correlation between CBS and CBV, r=0.960, P<0.001. CBS, clot burden score; CBV, clot burden volume; QD, Qanadli score.

Based on the automatic CBS and the manual Qanadli scoring, cases were divided into low-risk and high-risk groups (low <50%, high ≥50%). The average RVd/LVd ratio for the automatic scoring groups was 1.053±0.168 for the low-risk group and 1.118±0.223 for the high-risk group; for the manual scoring groups, the averages were 1.055±0.165 for the low-risk group and 1.123±0.249 for the high-risk group. There was no significant difference in the RVd/LVd ratio between the low-risk and high-risk groups (P=0.462 for CBS, P=0.627 for Qanadli scoring).


Discussion

This study has developed a high-performance VB-Net DL model for the automatic identification of PE and automatic evaluation of clot burden score based on CTPA images for the first time. The model performed excellently in random CTPA examinations, with a high PPV (92.00%) and a low false positive rate (0.21 per case); the CBS and CBV calculated automatically by the model were correlated with the manually measured Qanadli scores, and also showed the same trend with the increase in right heart parameters.

Accurately identifying CTPA images is fundamental to the diagnosis of APE. In recent years, research on the diagnosis of PE in CTPA using artificial intelligence (AI) has made phased progress (15). Some existing studies trained models with a small number of cases, or had a higher rate of false positives (16,17). Our model, after repeated optimization, had greatly reduced the number of false positives while trying not to affect the detection rate of clots above the subsegmental level. At the same time, the PPV, a key indicator for clinical application (18), was relatively high, which can reduce the rate of misdiagnosis and will not add extra medical burden. To explore the efficacy of the VB-Net DL model in actual clinical applications, this study randomly selected cases of APE. Due to the poor basic conditions and cooperation of some patients, image quality is often uncontrollable; among the 30 positive cases, seven cases were assessed as third-grade quality, which could easily have led to abnormal identification by the DL model, but it still demonstrated a high level of detection. There were seven cases of PE that the DL model failed to detect, which were single (6 cases) or double (1 case) subsegmental PA clots. The clinical value of single subsegmental clots is still controversial, and the outcomes with or without anticoagulation treatment may not differ. The 2019 guidelines of ESC and ERS recommend deciding whether to initiate anticoagulation treatment based on clinical conditions and bleeding risk (1,19). There were two false positive cases due to misidentification of main trunk clots, one due to interference from superior vena cava contrast artifacts and the other due to lung cancer and compression of the PA by enlarged lymph nodes. Uneven contrast filling, breathing artifacts, and cases after partial lobectomy or segmentectomy were also prone to misidentification. These interferences still pose a significant challenge to the model’s automatic recognition and further adjustments and optimizations are needed to improve the accuracy of the model in vascular boundary delineation.

Clot burden is related to the prognosis of patients with APE. When the obstruction of the pulmonary vascular tree reaches 30–50%, it can cause pulmonary hypertension, leading to a poor prognosis of PE (7); hence, it is necessary to quantitatively assess the obstruction of pulmonary vessels in APE patients (7,20). At present, the Qanadli score is one of the most frequently cited indicators in the literature. It is a semi-quantitative parameter that reflects the proportion of obstruction of the pulmonary vascular tree. The score is directly proportional to the severity of PE with high repeatability.

Previous studies have used DL models to automatically measure the clot burden volume (9,17), but the score reflected the obstruction ratio of the pulmonary vascular tree better than the clot volume. This study developed for the first time a high-quality VB-Net DL model to automatically measure the CBS and CBV in APE patients. The results showed that the CBS and CBV measured automatically by the DL model were highly correlated with the manually measured Qanadli scores, indicating that our VB-Net model’s measurement results were valid and reliable. In addition, the manual calculation of the Qanadli score is complex and tedious, and depends on the radiologist’s work experience, whereas our model can automatically calculate the clot volume and CBS, greatly saving the time of radiologists. Therefore, automatic scoring is more effective and convenient, which helps in the assessment of PE.

The relationship between clot burden and other clinical indicators is also very important, such as indicators used for PE risk stratification, right ventricular dysfunction. It is a parameter that many studies have asserted can predict prognosis (10,21), and there are several methods that can be used to measure it. Furlan et al. (5) concluded that an increased RVd to LVd ratio (P=0.001) is the only independent factor associated with short-term mortality. Meinel et al. (10) also found similar results in their meta-analysis that an increased RVd to LVd ratio measured on transverse CT images is associated with the highest risk, with a 2.5-fold increase in all-cause mortality and a 5-fold increase in the risk of death related to PE. The degree of pulmonary vascular obstruction is considered the most important factor in determining the right ventricle’s response to APE (22). Abdelwahab et al. (23) found a significant correlation between the volume of the clot and the parameters of right ventricular dysfunction assessed by CTPA, but no significant correlation between the two when using echocardiography. The results of Foley et al. (24) indicated that CTPA-based RVd/LVd ratio can be reliably automatically measured in most real cases of acute PE with perfect repeatability. The study by Zhang et al. (17) also confirmed this. In this study, the RV/LV ratio of the high-risk group, measured automatically by the model for clot burden scoring, showed an upward trend compared to the low-risk group, which is consistent with the conclusions of previous studies (8,9,17). However, the difference between groups in this study was not statistically significant, possibly due to insufficient data volume, and further comparative studies are needed.

There are still some limitations in this study. Firstly, we did not have a sufficiently large sample size to test the model’s reliability, but all data in this study were randomly selected, which makes the evidence more credible, and further supplemented the examination of the DL model’s efficacy from the perspective of the number of clots. In future studies, we will further increase the data volume for assessment and further explore the clinical application value of CTPA data. Secondly, the model’s ability to differentiate chronic PE still needs to be investigated. Thirdly, we only compared the potential relationship between clot burden and imaging parameters, without studying the relationship with other clinical parameters, and the correlation with clinical risk is not yet clear enough. Further assessment is needed regarding the correlation with patient risk stratification and predictive value for prognosis, which would help in the stratified management of patients with PE.


Conclusions

Our study demonstrates that the VB-Net model displayed good efficacy in the random data from our center, with advantages of being efficient and accurate in assessment, and has good clinical applicability. The automatic scoring of PE is expected to become an auxiliary tool for improving clinical diagnostic and treatment decisions, helping to accelerate the diagnosis and treatment process for critical cases and better risk stratification.


Acknowledgments

None.


Footnote

Reporting Checklist: The authors have completed the TRIPOD + AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-24-1412/rc

Funding: This work was supported by the Joint Research Development Project between Shenkang and United Imaging on Clinical Research and Translation (No. SKLY2022CRT201 to Y.Q., C.Y., X.Y., and M.Z).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-24-1412/coif). Y.Q., C.Y., X.Y., and M.Z. report that this work was supported by the Joint Research Development Project between Shenkang and United Imaging on Clinical Research and Translation (No. SKLY2022CRT201). Y.G. and Y.C. report that they are employed by United Imaging Intelligence, a subsidiary of Shanghai United Imaging Healthcare Co., Ltd. The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This study was approved by the Institutional Review Board of Zhongshan Hospital Affiliated to Fudan University (No. B2021-555). The requirement for informed consent was waived due to the retrospective nature of the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.


References

  1. Konstantinides SV, Meyer G, Becattini C, Bueno H, Geersing GJ, Harjola VP, et al. 2019 ESC Guidelines for the diagnosis and management of acute pulmonary embolism developed in collaboration with the European Respiratory Society (ERS). Eur Heart J 2020;41:543-603. [Crossref] [PubMed]
  2. LeVarge BL, Wright CD, Rodriguez-Lopez JM. Surgical Management of Acute and Chronic Pulmonary Embolism. Clin Chest Med 2018;39:659-67. [Crossref] [PubMed]
  3. Smith SB, Geske JB, Maguire JM, Zane NA, Carter RE, Morgenthaler TI. Early anticoagulation is associated with reduced mortality for acute pulmonary embolism. Chest 2010;137:1382-90. [Crossref] [PubMed]
  4. Shen C, Yu N, Wen L, Zhou S, Dong F, Liu M, Guo Y. Risk stratification of acute pulmonary embolism based on the clot volume and right ventricular dysfunction on CT pulmonary angiography. Clin Respir J 2019;13:674-82. [Crossref] [PubMed]
  5. Furlan A, Aghayev A, Chang CC, Patil A, Jeon KN, Park B, Fetzer DT, Saul M, Roberts MS, Bae KT. Short-term mortality in acute pulmonary embolism: clot burden and signs of right heart dysfunction at CT pulmonary angiography. Radiology 2012;265:283-93. [Crossref] [PubMed]
  6. Ouriel K, Ouriel RL, Lim YJ, Piazza G, Goldhaber SZ. Computed tomography angiography with pulmonary artery thrombus burden and right-to-left ventricular diameter ratio after pulmonary embolism. Vascular 2017;25:54-62. [Crossref] [PubMed]
  7. Qanadli SD, El Hajjam M, Vieillard-Baron A, Joseph T, Mesurolle B, Oliva VL, Barré O, Bruckert F, Dubourg O, Lacombe P. New CT index to quantify arterial obstruction in pulmonary embolism: comparison with angiographic index and echocardiography. AJR Am J Roentgenol 2001;176:1415-20. [Crossref] [PubMed]
  8. Fink MA, Mayer VL, Schneider T, Seibold C, Stiefelhagen R, Kleesiek J, Weber TF, Kauczor HU. CT Angiography Clot Burden Score from Data Mining of Structured Reports for Pulmonary Embolism. Radiology 2022;302:175-84. [Crossref] [PubMed]
  9. Liu W, Liu M, Guo X, Zhang P, Zhang L, Zhang R, Kang H, Zhai Z, Tao X, Wan J, Xie S. Evaluation of acute pulmonary embolism and clot burden on CTPA with deep learning. Eur Radiol 2020;30:3567-75. [Crossref] [PubMed]
  10. Meinel FG, Nance JW Jr, Schoepf UJ, Hoffmann VS, Thierfelder KM, Costello P, Goldhaber SZ, Bamberg F. Predictive Value of Computed Tomography in Acute Pulmonary Embolism: Systematic Review and Meta-analysis. Am J Med 2015;128:747-59.e2. [Crossref] [PubMed]
  11. Han M, Yao G, Zhang W, Mu G, Zhan Y, Zhou X, Gao Y. Segmentation of CT thoracic organs by multi-resolution VB-nets. In: International Symposium on Biomedical Imaging (ISBI) 2019.Available online: https://pagesperso.litislab.fr/cpetitjean/wp-content/uploads/sites/19/2019/04/SegTHOR2019_paper_1.pdf
  12. Shi F, Hu W, Wu J, Han M, Wang J, Zhang W, Zhou Q, Zhou J, Wei Y, Shao Y, Chen Y, Yu Y, Cao X, Zhan Y, Zhou XS, Gao Y, Shen D. Deep learning empowered volume delineation of whole-body organs-at-risk for accelerated radiotherapy. Nat Commun 2022;13:6566. [Crossref] [PubMed]
  13. Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser L, Polosukhin I. Attention is all you need. In: Advances in Neural Information Processing Systems (NeurIPS) 2017;30.
  14. Wittram C, Maher MM, Yoo AJ, Kalra MK, Shepard JA, McLoud TC. CT angiography of pulmonary embolism: diagnostic criteria and causes of misdiagnosis. Radiographics 2004;24:1219-38. [Crossref] [PubMed]
  15. Weikert T, Winkel DJ, Bremerich J, Stieltjes B, Parmar V, Sauter AW, Sommer G. Automated detection of pulmonary embolism in CT pulmonary angiograms using an AI-powered algorithm. Eur Radiol 2020;30:6545-53. [Crossref] [PubMed]
  16. Li X, Wang X, Yang X, Lin Y, Huang Z. Preliminary study on artificial intelligence diagnosis of pulmonary embolism based on computer in-depth study. Ann Transl Med 2021;9:838. [Crossref] [PubMed]
  17. Zhang H, Cheng Y, Chen Z, Cong X, Kang H, Zhang R, Guo X, Liu M. Clot burden of acute pulmonary thromboembolism: comparison of two deep learning algorithms, Qanadli score, and Mastora score. Quant Imaging Med Surg 2022;12:66-79. [Crossref] [PubMed]
  18. Altman DG, Bland JM. Diagnostic tests 2: Predictive Values. BMJ 1994;309:102. [Crossref] [PubMed]
  19. Bariteau A, Stewart LK, Emmett TW, Kline JA. Systematic Review and Meta-analysis of Outcomes of Patients With Subsegmental Pulmonary Embolism With and Without Anticoagulation Treatment. Acad Emerg Med 2018;25:828-35. [Crossref] [PubMed]
  20. Moroni AL, Bosson JL, Hohn N, Carpentier F, Pernod G, Ferretti GR. Non-severe pulmonary embolism: prognostic CT findings. Eur J Radiol 2011;79:452-8. [Crossref] [PubMed]
  21. Beenen LFM, Bossuyt PMM, Stoker J, Middeldorp S. Prognostic value of cardiovascular parameters in computed tomography pulmonary angiography in patients with acute pulmonary embolism. Eur Respir J 2018;52:1702611. [Crossref] [PubMed]
  22. Lualdi JC, Goldhaber SZ. Right ventricular dysfunction after acute pulmonary embolism: pathophysiologic factors, detection, and therapeutic implications. Am Heart J 1995;130:1276-82. [Crossref] [PubMed]
  23. Abdelwahab HW, Arafa S, Bondok K, Batouty N, Bakeer M. Relationship between clot burden in pulmonary computed tomography angiography and different parameters of right cardiac dysfunction in acute pulmonary embolism. Cardiovasc J Afr 2020;31:21-4. [Crossref] [PubMed]
  24. Foley RW, Glenn-Cox S, Rossdale J, Mynott G, Burnett TA, Brown WJH, Peter E, Hudson BJ, Ross RVM, Suntharalingam J, Robinson G, Rodrigues JCL. Automated calculation of the right ventricle to left ventricle ratio on CT for the risk stratification of patients with acute pulmonary embolism. Eur Radiol 2021;31:6013-20. [Crossref] [PubMed]
Cite this article as: Qiao Y, Gao Y, Chen Y, Ye X, Yan C, Zeng M. Quantitative assessment and risk stratification of random acute pulmonary embolism cases using a deep learning model based on computed tomography pulmonary angiography images. Quant Imaging Med Surg 2025;15(3):1950-1962. doi: 10.21037/qims-24-1412

Download Citation