Deep learning-based identification of chronic pulmonary embolism on CTPA: a regional lung analysis using multiplanar MIP images.

The CNN model achieved good performance in distinguishing CPE from non-PE and APE cases using full lung volume MIP images.

• Cross-validation AUROC of 0.80 was achieved for CPE classification using full lung volume MIPs
• CPE was classified against a combined APE and non-PE group
• The study included 41 CPE, 41 APE, and 41 normal controls (non-PE) for a total of 123 patients
• Performance decreased with reduced data as fewer vascular regions were included

The CNN model achieved higher performance for CTEPH classification compared to CPE classification.

• AUROC of 0.88 was achieved for CTEPH classification using full lung volume data
• 25 of the 41 CPE patients had CTEPH confirmed by right heart catheterization
• CTEPH was classified against a combined non-PE and APE group
• AUROC of 0.86 was achieved using only the most proximal half of the lung volume for CTEPH classification

Proximal vessels were identified as the most diagnostically relevant region for CTEPH detection.

• Using only the most proximal half of the lung volume yielded AUROC of 0.86 for CTEPH, close to the full-data AUROC of 0.88
• Central layers were most important for identifying CTEPH features
• CTPA data were divided into four concentric anatomic layers for regional analysis
• Performance for CPE detection was more dependent on inclusion of the full lung volume compared to CTEPH detection

Excluding proximal vessels by using an open-source segmentation model resulted in lower AUROCs for both CPE and CTEPH classification.

• The open-source segmentation model, which excludes proximal vessels, resulted in AUROC of 0.74 for CPE classification
• The open-source segmentation model resulted in AUROC of 0.83 for CTEPH classification
• These values were lower than the corresponding AUROCs of 0.80 (CPE) and 0.88 (CTEPH) achieved with the full vessel approach
• This comparison underscores the diagnostic importance of proximal vessel inclusion in the segmentation approach

Eleven masking schemes were applied to both classification tasks to assess the diagnostic value of different vascular regions.

• A total of 22 experiments were conducted (11 masking schemes × 2 classification tasks)
• The masking schemes varied the inclusion of different anatomic layers of the lung volume
• Model performances were compared using areas under the receiver operating characteristic curves (AUROC)
• The novel approach included proximal pulmonary vessels and a layered segmentation of the lung volume

Chronic pulmonary embolism and CTEPH are diagnostically challenging conditions associated with increased mortality when detection is delayed.

• CPE and CTEPH are described as 'challenging to diagnose' with delayed detection increasing mortality
• The study used CTPA-derived maximum intensity projection (MIP) images as input for the CNN
• CTEPH diagnosis was confirmed by right heart catheterization in 25 of 41 CPE patients
• The study aimed to evaluate CNN performance as a potential tool to improve future imaging diagnostics