RPS 1404b - Latest techniques in imaging of pulmonary vascular disease

Purpose:

To assess machine learning-based cardiac chamber segmentation in computed tomography pulmonary angiographies (CTPA) for the noninvasive detection of pulmonary hypertension (PH).

Methods and materials:

56 consecutive patients undergoing right heart catheterisation (RHC) with a measurement of mean pulmonary arterial pressure (mPAP) and CTPA for suspected PH within 3 weeks between 08/2013 and 09/2014 were retrospectively reviewed. Cardiac chamber segmentation was performed automatically by a machine learning-based commercially available software and manually corrected if deemed necessary. Cardiac chamber volumes (right and left ventricles (RV/LV) and right and left atria (RA/LA)) were indexed to body surface area.

Results:

37 patients were diagnosed with PH (mean mPAP 38±11 mmHg) and 19 patients had a normal mPAP (mean 17±4 mmHg). Manual correction of the segmentation was performed in 15 cases (27%). In patients with PH, RA, and RV were significantly enlarged (87±45 vs. 44±19 ml/m² and 102±39 vs. 63±23 ml/m²) and right to left atrial and ventricular ratios significantly increased (1.8±0.9 vs. 1.1±0.4 and 2.7±1.4 vs. 1.4±0.4) (all p<0.001). AUCs for the detection of PH were 0.83, 0.85, 0.76, and 0.83 for RA, RV, and atrial and ventricular ratio, respectively. A cutoff value for RV of 63 ml/m² allowed the diagnosis of PH with sensitivity, specificity, and positive and negative predictive values of 87%, 74%, 86%, and 74%, respectively.

Conclusion:

Machine learning-based automated cardiac chamber segmentation in CTPA allowed noninvasive detection of PH as confirmed by RHC with good diagnostic accuracy.

Limitations:

A retrospective and single-centre study with a limited number of patients.

Ethics committee approval

Ethics committee approved study and waived written informed consent.

Funding:

License for the segmentation software provided by Philips.

Purpose:

Pulmonary embolism (PE) is a life-threatening condition. Rapid and accurate risk stratification can decrease PE mortality rates. We aimed to develop a machine learning 30-day mortality predictive model using clinical data.

Methods and materials:

This retrospective study was approved by our institutional review board. We retrieved data for consecutive patients who underwent CT-angiography between 1/2012 to 12/2018 in our emergency department (ED) and were diagnosed with PE. Clinical variables included demographics, vital signs, chief complaint, medical history, home medications, blood tests, previous ED visits, and hospitalisations. The solution applies a machine learning-based classification algorithm, specifically, gradient boosting decision tree model (CatBoost). The model was trained on 2012-2017 data and tested on 2018 data. We evaluated the AUCs of single variables and of the entire model. We used Youden’s index to find the model’s optimal sensitivity and specificity.

Results:

Our dataset included 367 patients, 38 of whom (10.3%) died within 30 days of diagnosis. Single variables with higher AUC included pulse (0.72), Di-Dimer (0.71), and arrival by ambulance (0.71). The entire model achieved an AUC of 0.84 (95% CI: 0.757-0.926) for predicting PE severity. 78.4% of the test cases were correctly risk stratified, based on the 30-day mortality label. Using Youden’s index, the model showed a sensitivity of 90.9% and specificity of 77.7% for predicting death within 30 days.

Conclusion:

Our model allows effective risk stratification of individuals with PE, which could improve the ability to direct preventative and health surveillance resources. To the best of our knowledge, the system is the first such fully automated solution.

Limitations:

A retrospective single-centre study with a relatively small number of patients.

Ethics committee approval

Approved by our institutional review board.

Funding:

No funding was received for this work.

Purpose:

To explore the potential improvement in pulmonary artery opacification and to assess the change in image quality parameters in VMI using fast switch kVp dual-energy CT.

Methods and materials:

Computed tomography angiography (CTA) images of 877 patients with a diagnosis of PE were reviewed. 60 patients with suboptimal enhancement (<200 HU) were involved. Standard images (140 kVp) and VMI from 40 to120 keV were generated. Attenuation, noise, SNR, and CNR were measured in the pulmonary artery (PA). Using the VMIs, the best image was determined as the image with main PA opacification greater than 200 HU and image quality ≥3. 56 studies that met these criteria were considered as salvaged. At this best energy level, quantitative parameters were compared with standard images.

Results:

The attenuations of VMIs at 40, 45, 50, 55, 60, 65, and 70 keV were significantly higher than standard images (p< 0.001). Similar findings were observed with SNR and CNR. In the salvaged patients, the average increase in mean PA attenuation was 62% (from 172.61±23.4 to 280.55±40.7), the average increase in SNR was 38% (from12.1±5.3 to 16.7±7.1), and the average increase in CNR was 48% (9.2±4.3 to 13.7±6) (p< 0.001).

Conclusion:

Low keV VMI reconstructions significantly increase PA attenuation, CNR, and SNR compared to standard image reconstructions. Suboptimal CT studies can be salvaged using low keV VMIs.

Limitations:

The study design is retrospective and includes a single-centre experience. To determine the adequacy of PA opacification level in CTA, we accepted a limit of 200 HU. There is no single predetermined standard value that can be used in this regard for pulmonary CTA.

Ethics committee approval

Approved by the Institutional Review Board(Approval Number:19-3.1T/70).

Funding:

No funding was received for this work.