Diagnoses of pulmonary embolism from non-contrast 4DCT using image processing-derived quantitative perfusion scores

Abstract

Computed tomography pulmonary angiography (CTPA) is the gold standard for pulmonary embolism (PE) diagnosis, but patients with iodinated contrast allergies or renal insufficiency are often ineligible. CT-derived perfusion (CTP) is a novel, non-contrast method to quantify pulmonary perfusion from an inhale/exhale CT image pair (4DCT). The resulting CT-P information can be used to identify hypo-perfused regions associated with PE. This pilot study introduces a thresholding approach that estimates the number of lung lobes with perfusion deficits according to optimally selected CTP thresholds. The number of lobes indicated as low-functioning provides a score to categorize patients as PE-positive, negative, or inconclusive. We trained and validated the model on a retrospective dataset of 123 suspected PE patients, achieving 72% accuracy, 75% sensitivity, and 69% specificity, with 17% of cases inconclusive. To our knowledge, this is the first PE diagnostic model from non-contrast 4DCT, showing the feasibility of non-contrast PE diagnosis strategies.

Introduction

Pulmonary embolism (PE) is a life-threatening condition in which an embolus blocks blood flow to the lungs, causing 300,000 deaths yearly in the United States^1,2. Untreated PE has been associated with a high short-term mortality rate of up to 33%, whereas the death rate of diagnosed and treated PE is only 8%, indicating a pressing need for early diagnosis and treatment^3,4. PE is difficult to diagnose due to its non-specific signs and symptoms, which overlap with conditions such as pneumonia, myocardial infarction, and pleuritis. Of patients with PE, 27.5% are misdiagnosed initially in emergency department settings, and 53.6% of patients are misdiagnosed in inpatient settings⁴. Currently, there are two validated imaging modalities for PE diagnosis: planar ventilation/perfusion (V/Q) scan and CT pulmonary angiography (CTPA). The V/Q scan marked the first attempt to replace pulmonary angiography, but results must be interpreted in combination with other tests, which are often perceived as too complex by clinicians in high patient volume areas⁵.

Besides V/Q scans, other alternative diagnostic tests include contrast-enhanced magnetic resonance angiography (CE-MRA), non-contrast MRA, and single photon emission computed tomography (SPECT)^6,7. While MRA can diagnose PE quite accurately, it is also more technically demanding than CTPA. MRA requires longer examination times and is only advised in well-experienced facilities for patients with contraindications to standard tests⁸. Holistic SPECT ventilation/perfusion diagnosis strategies are based on identifying segmental or subsegmental functional mismatches between ventilation and perfusion. Holistic interpretation of SPECT perfusion (typically acquired with a 99mTc-labeled macro-aggregates of albumin 99mTcMAA tracer⁶) and Technegas ventilation scans have been demonstrated to diagnose PE with 98% sensitivity and 99% specificity⁸. In contrast to CTPA, SPECT imaging can be safely acquired for 99% of patients with a sensitivity of 0.96^9–11, but requires transportation of the patient to a nuclear medicine clinic. As a result, holistic nuclear medicine PE diagnosis requires a longer scan acquisition time (1–2 h) and may not be available on nights and weekends at many hospitals. As delays in PE diagnosis are critical and problematic¹², a rapid method for PE diagnosis based on non-contrast imaging is necessary.

Due to its simplicity, accuracy, and high accessibility, CTPA is considered the current gold standard for the diagnosis of PE¹³.CTPA provides a highly accurate and rapid diagnosis with a sensitivity of 90% and specificity of 95% when combined with CT venography^13,14. CTPA however, has some risks and contraindications, such as allergies to contrast and renal failure¹⁴. The high radiation dose of CTPA to the chest is also an increasing concern, particularly in young women. Studies suggest that a woman undergoing a CTPA at the age of 20 could have a 0.5% absolute increase in lifetime risk of breast cancer¹⁵.

To address the unmet medical need for diagnosing pulmonary embolism in patients who are allergic to contrast agents, several studies have explored the feasibility of using non-contrast CT scans for PE diagnosis^16–19. In 2019, Nafisa et al. evaluated the role of non-contrast CT in detecting PE in 100 patients and reported an overall sensitivity of 50% and specificity of 98.6% using the hyperdense lumen sign identified in unenhanced multidetector computed tomography (MDCT)²⁰. In the same year, Chien et al. reported a sensitivity of 96.9% and specificity of 71.9% using unenhanced MDCT in a cohort of 32 positive and 32 negative cases²¹. In addition to the hyperdense lumen sign, pulmonary infarction has also been used as an indicator to detect PE from non-contrast CT²². These studies demonstrate that non-contrast chest CT is an alternative for the diagnosis of acute PE when CTPA is not available, showing the feasibility of diagnosing PE from non-contrast CT.

However, while previous studies have shown correlations between CTPA and non-contrast CT, the MDCT-based approaches proposed in their study mostly focus on the hyperdense lumen sign, with case classifications performed by radiologists. Despite its usefulness for PE detection, the hyperdense lumen sign is not always visible in non-contrast CT, making it challenging for clinicians to identify reliably and without bias. This explanation might explain the large gaps in sensitivity and specificity reported in previous studies (50% with 98.6% and 96.9% with 71.9%). Building upon these non-contrast PE detection approaches, we developed a mathematical method to diagnose PE based on pulmonary perfusion, a direct physiological indicator of PE, with the goal of minimizing bias and achieving both high sensitivity and specificity using non-contrast CT.

Pulmonary perfusion refers to blood flow to alveolar capillaries. PE occurs when a blood clot occludes the pulmonary artery, reducing perfusion distal to the clot, which makes low perfusion a key indicator of PE. In previous studies, we developed image-processing-based approaches to quantify pulmonary ventilation and perfusion from dynamic non-contrast CT scans (4DCTs), which we refer to as CT-derived function imaging (CTFI)^23–25. Unlike CTPA, four-dimensional computed tomography (4DCT), a series of 3D lung CT scans taken during respiration, does not involve injecting any contrast media. 4DCTs are standard 3D CT images resolved into different phases of the breathing cycle⁹. 4DCT has been routinely utilized in radiation oncology to provide dynamic imaging throughout the respiratory cycle and can be acquired using standard CT scanners available in Emergency Centers. Thus, PE diagnostics based on 4DCT would have the benefit of being accessible and rapid.

While CT-derived ventilation imaging has been well-studied^26,27, CT-derived perfusion is a newer concept. In previous work, we have shown that CT-perfusion (CT-P) has a good correlation with SPECT-perfusion²⁴ and that the methodology has the potential to be a tool for diagnosing pulmonary embolism from non-contrast CT scans²⁸. CT-P quantifies subtle spatial pulmonary-blood mass distribution changes between inhale and exhale CT scans as a surrogate measure for blood flow. We showed that pulmonary blood mass change is lower in PE-positive patients, suggesting the feasibility of CT-P-based diagnosis. While CTPA diagnosis relies on radiologist interpretation, CT-P provides quantitative values to define criteria for identifying PE patients.

In this study, we present a novel CT-P-based thresholding strategy for identifying patients with PE from non-contrast 4DCT images. We aim to demonstrate the feasibility of diagnosing PE using non-contrast CT, and our proposed model represents the first quantitative diagnostic score for detecting pulmonary embolism from non-contrast CT imaging. As the model is based on quantitative values, our approach provides a clinically explainable tool for pulmonary embolism diagnosis that can be acquired on existing emergency center scanners. Building upon the promising results we show in this pilot study, we prove the feasibility of diagnosing PE using non-contrast 4DCT.

Results

Data distribution

We analyzed the distribution of our CT-perfusion scores across all study patients. Perfusion scores for positive PE cases are significantly lower than negative cases in lobes 1, 2, 4, and 5 (p = 0.0057, 0.0005, 0.0022, and 0.0005, tested using two-sample t-test) as shown in Fig. 1.

Fig. 1. Boxplots of perfusion data for lobes 1 to 5.

These boxplots show that the medians of pulmonary perfusion in the 64 positive cases are lower than those in the 59 negative cases for each lobe. Statistical differences between the positive and negative groups were calculated using a two-sample t-test after the outliers were removed (11 cases for positive and negative, respectively). P-values for Lobe 1 to 5 in graphs a–e are 0.027, 0.003, 0.352, 0.023, 0.007. ***p < 0.001, **p < 0.01, *p < 0.05.

Model performance

Using LOOCV, our model achieved an accuracy = 0.72, a sensitivity = 0.75, and a specificity = 0.69 with 17% inconclusive cases. Including inconclusive cases, accuracy dropped to 0.66, sensitivity to 0.60, and specificity to 0.73 (Table 1). The results showed that allowing inconclusive cases improved accuracy. Table 2 shows the average and standard deviation of threshold values for each lobe over LOOCV runs.

Table 1.

Model performance.

	Sensitivity	Specificity	Accuracy	Inc (%)
W/o Inc	0.60	0.73	0.66	0
With Inc	0.75	0.69	0.72	17

Sensitivity, specificity, and accuracy were evaluated on the test data as mentioned in “Leave-one-out cross-validation and model evaluation”. The inconclusive rate was calculated on the training data during the optimization process. Our model achieved a sensitivity of 0.75 and a specificity of 0.69, with 17% of cases being inconclusive (Inc). Without any inconclusive cases, it achieved a sensitivity of 0.60 and a specificity of 0.73.

Table 2.

Average (Avg) and standard deviation (Stdev) of threshold values for each lobe over cross-validation runs.

	Lobe 1	Lobe 2	Lobe 3	Lobe 4	lobe 5
Avg	2.75E−03	1.23E−02	1.78E−03	9.93E−03	6.50E−03
Stdev	0.0041	0.0009	0.0005	0.0046	0.0008

The standard deviation for each lobe is lower than 0.005, validating the robustness of the thresholds found using our optimization approach.

Impact of each lobe

Our PE score is the number of lobes considered to have a perfusion deficit according to threshold values. We also investigated whether the binary signature, which indicates the lobes in deficit for a patient, affects the diagnostic prediction. Since the binary signature is 5 digits long, there are a total of 2⁵ = 32 possible signatures. The percentage of times each signature appeared within the positive and negative LOOCV testing is shown in Fig. 2. For simplicity, we denote binary signatures as the lobes in deficit. For instance, Lobe 24 refers to signature [01010].

Fig. 2. The percentage of each binary signature taken up of positive/negative cases.

S denotes the diagnostic score. The lobe numbers were defined as labeled in the lung figure, where the three right lobes were labeled lobe 1 to 3 (from top to bottom), and the two left lobes were labeled 4 to 5 (top to bottom). The percentage of each diagnostic score taken up in each category is also demonstrated in the table. 3.1% of positive and 22% of negative cases were taken up by cases with S = 0, 17.1% of positive and 37.3% negative cases for S = 1, 20.3% of positive and 13.6% negative cases for S = 2, 39.1% of positive and 18.6% of negative cases for S = 3, 20.3% of positive and 8.5% of negative cases for S = 4, and 0% for both positive and negative cases for S = 5.

Several signatures were observed to only occur in positive cases, such as Lobe 12, 124, 1235, and 1245. Similarly, some only occurred in negative cases, such as Lobe 3, 15, 34, and 235. Signatures of many negative cases include lobe 3, which is the only lobe that does not show a significant difference between positive and negative cases (Fig. 1).

Method

Model overview

We devised a scoring scheme based on estimating the number of lung lobes experiencing a perfusion deficit according to optimal CTP thresholds. Our pipeline for extracting CT-P values from 4DCT and determining optimal threshold values is detailed in Fig. 3. For data preprocessing (marked gray in Fig. 3), we first convert the CT Hounsfield Units into estimates of material density values following the derivation in refs. ^26,28. To calculate the pulmonary blood mass change, we apply automated DenseNET segmentation to delineate lung lobe volumes on the inhale and exhale scans²⁹. The magnitude of mass change per lobe is calculated as the absolute value of the difference between inhale and exhale lobe masses. For classification (marked blue in Fig. 3), we optimize the perfusion deficit vs. normal-functioning threshold for each lobe. These optimized thresholds are used to binarize the lobar CT-P data with 1’s assigned to lobes in perfusion deficit, and 0’s for those with functional perfusion (values higher than the threshold). To calculate the final diagnostic score based on the perfusion deficit in each lobe, the binarized scores are summed up to arrive at a diagnostic score ranging from 0 to 5 for each patient. The scoring system was developed based on the assumption that low perfusion in each lobe carries a similar weight as an indicator of PE. Finally, diagnostic predictions are made based on scores: < 2 are PE-negative, > 2 are PE-positive, and = 2 are inconclusive.

Fig. 3. Overview of the model.

Our proposed approach includes the data-preprocessing and classification parts. L1-L5 denotes lobe 1 to lobe 5 as labeled in the top-right lung figure, and case 14 was used as an example. We extracted the pulmonary perfusion data from the 4DCT images by converting the CT numbers into density and further calculating the mass change in each lobe using the density. To classify the cases into positive and negative, we found the threshold for pulmonary perfusion in each lobe and used the thresholds to convert mass change data into a 1 × 5 binary matrix for each case. By summing up the binary matrix, we get the diagnostic score. Finally, we classified cases with a score > 2 as positive, <2 as negative, and = 2 as inconclusive. Equations listed in the figure will be further discussed in detail in “Methods“.

Patient data

We conducted an Institutional Review Board-approved imaging trial with 129 patients who presented to the Emergency Center at our institution (William Beaumont University Hospital, Royal Oak, MI) with suspected pulmonary emboli (Clinical Trials Registration: NCT03183063). Each patient received a CTPA for first-time acute PE diagnosis and a non-contrast 4DCT scan within 48 h of the CTPA acquisition. 4DCT images were acquired on a General Electric Revolution Evo 64-slice scanner operating in sequential cine mode. Inhale and exhale phase images were reconstructed from cine data using a markerless phase binning algorithm²⁷. The resulting image phases possessed 0.97 × 0.97 mm axial-plane voxel spacing and 2.5 mm slice thickness. Of the 129 total study patients, 60 were PE-negative and 69 PE-positive according to the gold standard CTPA. A total of 123 cases were used in this study, including 59 negative and 64 positive cases, after removing 6 cases where the DenseNET²⁹ lobe segmentation failed, primarily due to significant image acquisition artifacts and infarcted lung (details provided in the “Discussion” section, Fig. 5).

Fig. 5. PE cases with significant artifacts.

a and b: Failure example 1: The lung is infarcted. Lobes are not found. c and d: Failure example 2: Phase binning artifact erroneously includes diaphragm in lobe segmentation.

Mass change calculations

Physical density for each voxel is determined by converting Hounsfield units (HU) into material density estimates (g/mm³) under the assumption that each voxel represents a linear combination of air and tissue components:

$$\rho =1+\frac{\mathit{HU}}{1000}$$ 1

The HU scale is linear between air and water densities, where −1000 HU corresponds to air, and 0 HU corresponds to water. Inhale and exhale ROIs, including lobes 1 through 5, were segmented and labeled automatically using a DenseNet convolutional neural network model³⁰.

The mass for each lung lobe ROI (Ω) was calculated as the sum of voxel masses within the ROI:

$$m\left(\mathrm{\Omega }\right)=\sum _{x_i\in \mathrm{\Omega }}\left({\rho }_i\times v_i\right),$$ 2

where ${\rho }_i$ denotes the density of the voxel location $x_i$ within the lung lobe ROI Ω and v_i is the voxel volume. Knowing that variations in the apparent lung mass on dynamic imaging is attributed to variations in the amount of blood mass within the lung throughout the breath cycle²⁵, we calculated the differences in lung mass between exhale and inhale. To take the size difference of each lobe into account, the mass difference is normalized by the exhale lung lobe ROI ${\Omega }_{\mathit{ex}}$. Given the inhale lobe ROI ${\Omega }_{\mathit{in}}$ and exhale lung lobe ROI ${\Omega }_{\mathit{ex}}$, mass change for each lobe is calculates as:

$$M_l^n=\frac{\left|m\left({\Omega }_{\mathit{in}}\right)-m\left({\Omega }_{\mathit{ex}}\right)\right|}{{\Omega }_{\mathit{ex}}},$$ 3

where n is the patient case number and l ∈ [1,5] specifies the ROI (lung lobe). As the density of lung remains the same throughout one breathing cycle, we can assume that the observed change in density is solely caused by the blood flow and thus can use the calculated lung mass change between inhale and exhale, $M_l^n$, as a surrogate for lung perfusion.

Threshold optimization and diagnostic score calculation

To decide what differences in lung mass between exhale and inhale indicate a perfusion defect, we search for optimal cut-off thresholds for each lobe using Particle Swarm Optimization. After calculating CT-P values, we optimized for thresholds that can best separate positive and negative cases based on the perfusion values. We simplified the decision-making process by assigning each case a 1 × 5 vector, for example, B = [1, 1, 1, 0, 0], where each entry represents whether the lobe is considered functional (0) or non-functional (1):

$$B_l^n=\left\{\begin{array}{c}1,M_l^n<T_l\\ 0,M_l^n\ge T_l\end{array}\right),$$ 4

where n is the patient case number, and l specifies the ROI (lung lobe). A score of ‘1’ represents a lobe with a functional deficit, and 0 indicates normal function. Particle Swarm Optimization, a heuristic optimization method that does not require differentiability of the loss function, was used to search for the optimal thresholds^31,32 as illustrated in Fig. 4.

Fig. 4. Flowchart of the particle swarm optimization algorithm.

Once the optimization starts, a population of random solutions (in this study, 5 thresholds for 5 lung lobes) will be generated and used to classify pulmonary perfusion data into high or low perfusion in each lobe for each case. Next, the results of the five lobes for each case will be summed up to generate a diagnostic prediction. The objective function will be used to evaluate the fitness of the solutions in this population, and if the value of the objective function for the validation set or the training set, whichever is greater, is smaller than −0.7, the optimization will stop, and the solutions in this population will be used as the optimal solutions.

The minimization problem associated with our model is the determination of five CT-P threshold values (one for each lobe) that indicate whether the corresponding lobe has a functional deficit.

Assuming that lobes are more likely to be in deficit within the PE-positive population, we define our objective function used to evaluate the fitness of a candidate threshold set as:

$$\mathit{Loss}\left(T\right)=-\left[\left(\frac{N_P^{\mathit{predicted}}}{N_P^{\mathit{labeled}}},\frac{N_N^{\mathit{predicted}}}{N_N^{\mathit{labeled}}}\right)-0.15\times \frac{S_P}{N_P}+0.15\times \frac{S_N}{N_N}\right],$$ 5

where $S_P$ is the sum of the prediction score for all positive cases and $S_N$ for negative cases. $N_P$ and $N_N$ are the number of positive and negative cases in the training dataset, respectively. To calculate $N_P^{\mathit{predicted}}$ and $N_N^{\mathit{predicted}}$, cases with prediction scores > 2 will be labeled as positive, with scores < 2 as negative, and those with scores = 2 will not be included. $\frac{S_P}{N_P}$ and $\frac{S_N}{N_N}$ are the mean prediction scores for positive and negative cases, ranging from 0 to 5. $\frac{N_P^{\mathit{predicted}}}{N_P^{\mathit{labeled}}}$ and $\frac{N_N^{\mathit{predicted}}}{N_N^{\mathit{labeled}}}$ are the sensitivity and specificity of the diagnosis corresponding to the threshold values T:

$$\mathit{Sensitivity}=\frac{N_P^{\mathit{predicted}}}{N_P^{\mathit{labeled}}}$$ 6

$$\mathit{Specificity}=\frac{N_N^{\mathit{predicted}}}{N_N^{\mathit{labeled}}},$$ 7

where $N_P^{\mathit{predicted}}$ is the number of positive cases that are predicted to be positive (true positive), $N_P^{\mathit{labeled}}$ is the number of positive cases. $N_N^{\mathit{predicted}}$ is the number of negative cases that are predicted to be negative (true negative), $N_N^{\mathit{labeled}}$ is the number of negative cases.

To maximize the sensitivity and specificity, our goal is to maximize $\frac{S_P}{N_P}$ to maximize the number of true positive predictions and minimize $\frac{S_N}{N_N}$ to maximize the number of true negative predictions in Eq. (5). Therefore, we added a minus sign before the second term and a positive sign before the third term.

Considering that in our model, the minimum value between $\frac{N_P^{\mathit{predicted}}}{N_P^{\mathit{labeled}}}$ (sensitivity) and $\frac{N_N^{\mathit{predicted}}}{N_N^{\mathit{labeled}}}$ (specificity) falls between 0.65 to 0.75, we multiplied $\frac{S_P}{N_P}$ and $\frac{S_N}{N_N}$ (the mean prediction scores for positive and negative cases, ranging from 0 to 5) by 0.15 to ensure that the three terms all fall within the range from 0 to 0.75 and have equal impact on the loss function. Constant 0.15 was chosen to make the greatest possible value of the second and third terms 0.75. By assigning the weights, we could search for thresholds for maximum accuracy with minimal sensitivity-specificity difference, ensuring accuracy for both positive and negative cases. Considering these factors, we determined the stopping criteria, theoretically within −0.65 to −0.75, of −0.7 for the optimization algorithm through trial and error.

The final diagnostic score was calculated by summing up the result in each lobe (Eq. (8)). Cases with a diagnostic score > 2 were classified as positive while < 2 were negative.

$${\mathit{DS}}_n=\sum _l^n\left(B_l^n\right)$$ 8

Inconclusive cases

Similar to planar VQ PE assessment, we allow our 4DCT-based diagnostic model to return an inconclusive classification, meaning that the model cannot determine if the case is positive or negative. In our model, cases with a diagnostic score = 2 were labeled as inconclusive.

$${Prediction}_n=\left\{\begin{array}{c}Positiveif{DS}_n>2\hfill \\ Inconclusiveif{DS}_n=2\hfill \\ Negativeif{DS}_n<2\hfill \end{array}\right)$$ 9

Leave-one-out cross-validation and model evaluation

Considering the limited data available, we conducted leave-one-out cross-validation (LOOCV) to test the model. LOOCV withholds one data point from the training process and uses it to evaluate the model’s ability to generalize to new data. In each iteration, one case (0.8% of the dataset) was taken out from the dataset and was used as the test data after the training process, while the rest (122 cases, 99.2% of the dataset) were used as training data. For model evaluation, we calculated the overall prediction accuracy, the inconclusive rate, sensitivity, and specificity.

Discussion

Many recent studies focus on integrating artificial intelligence (AI) into PE diagnosis and achieving high sensitivity and specificity³³. Most AI models for PE diagnosis are convolutional neural network (CNN)-based deep learning models that operate on CTPA; however, low interpretability and large data requirements limit clinical applicability. Deep learning models are data-intensive, meaning that the amount and quality of data used to train significantly affect the model’s performance and can lead to overfitting. Additionally, deep learning models lack interpretability, making it difficult to apply them in clinics for medical decision-making. Such methods provide no benefit to patients contraindicated for iodine-based contrast. Our 4DCT-based diagnostic model is a simple mathematical model that searches for CT-derived perfusion thresholds to classify PE and non-PE cases, making it more clinically meaningful than existing deep learning models while training on smaller datasets.

Due to the complexity of 4DCT reconstruction and the patient motions during image acquisition, we found severe image artifacts in several cases (Fig. 5), causing our AI segmentation model, DenseNet, to fail for these cases. Given the biased sampling population and the artifacts that could affect the quantification of 4DCT-derived perfusion, we can conclude that the dataset used in this study is challenging and relatively small. However, even with this difficult dataset, our model was able to achieve a sensitivity of 0.75 and a specificity of 0.69, with 17% of cases being inconclusive, showing the high robustness and adaptability of our model.

Our PE diagnostic model has several strengths. First, it minimizes model complexity, enabling accurate classification with a small dataset, which deep learning models cannot achieve. Second, it ensures model robustness, detecting subtle differences in lung perfusion between PE and non-PE cases even with biased datasets. Participants in this study were required to give informed consent, which led to potential bias in the sampling population, as patients who were able to sign consent and undergo 4DCT imaging were likely those with less severe symptoms. The lack of severe cases was demonstrated in Fig. 2, where there are no cases with a diagnostic score of 5. Had massive and submassive PE cases been included, they would likely have yielded scores greater than 2 and would have been classified as PE positive by our method. In fact, compared with the mild cases included in this study, more severe PE cases would likely have been easier to classify, suggesting that our method may perform better in real-world clinical settings.

The limitations of this study include, as mentioned above, the potentially biased sampling population due to the 4DCT acquisition process. Although it highlights the robustness and sensitivity of our proposed model, we plan to further validate it using animal models and a larger human 4DCT. Secondly, the lobar thresholding approach performed well for this difficult dataset, demonstrating its effectiveness in diagnosing acute PE. The diagnostic cutoff was determined by solving a mathematical optimization problem (see “Threshold optimization and diagnostic score calculation” for details), which is a simple yet robust approach that enables PE classification using this small and challenging dataset. However, this model has only been tested on 123 cases and may not be fully generalizable yet, as the lobar 4DCT-derived perfusion could have been affected by artifacts or imperfect segmentation. 4DCT acquisition (phase-binning) artifacts can distort lung geometry and erroneously assign HU values within the lung volume. These geometric distortions can lead to inaccuracies in the lung lobe segmentations and mass measurements that the CT-P calculation depends on. In such cases, the lobar CTP values are essentially corrupted by the artifact and do not represent the patient’s true physiological state, thereby making any downstream diagnostic assessments equally inaccurate. This limitation may be mitigated by using inhale/exhale breath-hold CT pairs, which do not require phase binning, rather than 4DCT as the non-contrast imaging modality. The primary goal of this pilot study was to demonstrate the feasibility of using non-contrast 4DCT to diagnose acute PE, and we plan to validate and refine the diagnostic cutoff (diagnostic score > 2 as PE-positive) and further investigate the lobar impact in future work.

Compared to other non-AI methods for detecting PE from non-contrast CT described in “Introduction”, our approach provides more clinically meaningful and unbiased predictions by identifying regions of low pulmonary perfusion, a direct indicator of embolic obstruction, and replacing human classification with an automated mathematical model. Our model achieved high sensitivity and specificity (0.75 and 0.69, respectively) without human intervention, ensuring acceptable false-positive and false-negative rates. By quantifying pulmonary perfusion (CT-P) from non-contrast CT and developing a mathematical model to classify PE based on CT-P, we advanced PE detection beyond existing methods and believe it could be beneficial in clinical practice, especially in emergency settings. Furthermore, we hope to localize PE from non-contrast CT using our proposed approach after validation in animal models, which will be part of our future work.

Considering the difficulty encountered while collecting data and reconstructing 4DCT images, we suggest using breath-hold CT instead of 4DCT as the next step for this study. In our model, we compared the mass changes between inhale and exhale CT using a 4DCT dataset acquired in a previous study²⁸, though we only needed the inhale and exhale scans. For future work, we can use breath-hold CT instead of dynamic CT scans. Breath-hold CT can avoid the artifacts caused by phase binning reconstruction while being more accessible in the clinic. Though the simplicity of our diagnostics model allows for ease of implementation, our optimal threshold parameters are specific to a free-breathing maneuver, meaning that if the inhale and exhale images were acquired with a forced breath-hold maneuver, the thresholds would be different. Hence, our next step will be to test breath-hold data and classify the data using our diagnostic model, which is expected to have higher accuracy. Using breath-hold data instead of 4DCT should enable faster and simpler data acquisition, allowing us to collect more data and enabling image acquisition from patients with more severe PE. In this study, we observed a higher likelihood of low perfusion in lobes 2 and 4 among positive cases, which we believe is due to the characteristics of our sampling population. With a larger and more evenly distributed dataset, we could improve the precision of our diagnostic model and better understand the relationship between the location of low pulmonary perfusion and the occurrence of PE.

Author contributions

All authors contributed to the study conception and design. Material preparation and data collection, and analysis were performed by G.N., D.T.-L., and C.S. The first draft of the manuscript was written by H.-T.K., and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Data availability

The generated dataset used in the current study is available from the corresponding author on reasonable request.

Code availability

The code used for this study is written in MATLAB and includes the code for optimizing threshold selection and the code for computing the PE diagnosis score given threshold values. These codes, including a driver script with implementation instructions, are openly hosted on GitHub (https://github.com/DMIC-Lab/PE-detection-code.git) under the MIT License.

Competing interests

Edward Castillo reports a relationship with 4D Medical that includes: funding grants and licensed patent #10932744. All other authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Competing Interests

The authors declare no competing interests.