Abstract 17842: 3D Visualization and Quantitative Assessment of the Pulmonary Arteries on CT Using Deep Learning Segmentation

Introduction: Assessment of the pulmonary arteries (PA) on CT scan is typically qualitative and requires review of hundreds of images. PA segmentation would enable creation of 3D reconstructions that visualize the entire pulmonary arterial tree in a single image, similar to catheter angiography, and also enable quantification of pathophysiologic changes. Manual segmentation is very time-consuming, but deep-learning (DL) can automate the process.

Hypothesis: A DL segmentation algorithm can be used to automatically generate 3D reconstructions of the PAs and quantify morphologic changes seen in Chronic Thromboembolic Pulmonary Hypertension (CTEPH) and Pulmonary Arterial Hypertension (PAH).

Methods: A DL algorithm based on a 3D U-Net architecture which segments PAs larger than 2 mm was trained and validated on a total of 250 contrast-enhanced CT chest. The study group included 15 CTEPH and 15 PAH patients seen consecutively in pulmonary hypertension (PH) clinic in 2022 with prior CT chest and right heart catheterization, and 10 additional patients with prior CT chest and no evidence of PH on same day echo (normal). After running the DL algorithm on CTs, 3D renderings were created from segmentations and the relative volume of the central, interlobar, and segmental/subsegmental PAs was derived.

Results: 3D renderings of PAH and CTEPH patients showed larger central and interlobar PAs compared with those of normal patients, whereas segmental and subsegmental PAs were smaller and pruned in CTEPH patients (Figure). Quantitatively, the volume of the central and interlobar PAs was higher in CTEPH and PAH compared to normal (p < 0.05). The combined volume of segmental/subsegmental PAs was smaller in CTEPH patients relative to normal and PAH patients (p < 0.005).

Conclusions: DL segmentation enables intuitive visualization and quantification of the morphologic changes to the PAs seen in conditions causing PH. This method may improve diagnosis and assessment of PH on CT.