Spectral CTPA-based multi-modal deep learning model for acute pulmonary embolism detection and segmentation.

Objectives

To develop a multi-modal deep learning (DL) model based on spectral CTPA for automated detection and segmentation of acute pulmonary embolism (APE).

Methods

In this retrospective study, 526 participants (126 with APE) who underwent spectral CTPA across three cohorts were included and assigned to training, validation, internal test (ITS), and external test (ETS) sets. Each participant had both CTPA and iodine density maps (IDMs), and lesions were manually annotated as ground truth. Baseline DL models were first developed using either CTPA or IDM in a single-modal setting, and subsequently extended to a multi-modal framework by fusing CTPA-IDM pairs. Model performance was evaluated using the Dice similarity coefficient (DSC) and area under the receiver operating characteristic curve (AUC), with statistical analysis performed using the Wilcoxon signed-rank test and DeLong test.

Results

For patient-level APE detection, the multi-modal model achieved a higher AUC of 0.947 than the single-modal models using CTPA (0.914, p = 0.66) or IDM (0.884, p = 0.17). For slice-level detection and lesion segmentation in APE patients, the multi-modal model also outperformed the others, yielding the highest AUCs (ITS: 0.955; ETS: 0.945; all p < 0.001) and DSCs (ITS: 0.743; ETS: 0.729; all p < 0.001). Notably, in peripheral embolism detection, recall improved from 49.53% to 63.55% in ITS and from 57.73% to 72.16% in ETS compared to the single-modal model.

Conclusion

This pilot study suggests that a spectral CTPA-based multi-modal DL model may improve slice-level APE detection, embolism segmentation, and peripheral PE recall by integrating CTPA and IDM. Further prospective studies are needed to validate its patient-level detection benefit over single-modal CTPA.