Multi-modal deep learning on spectral CTPA improves pulmonary embolism detection and segmentation

Multi-modal deep learning on spectral CTPA improved slice-level PE detection (AUC 0.955 internal, 0.945 external) and peripheral embolism recall (from 49.53% to 63.55% internal, 57.73% to 72.16% external) vs. single-modal CTPA.

• Multi-modal model using spectral CTPA and iodine density maps achieved slice-level AUCs of 0.955 (internal test) and 0.945 (external test) for PE detection, significantly outperforming single-modal CTPA (all p<0.001).
• Lesion segmentation DSC improved to 0.743 (internal) and 0.729 (external) (p<0.001); peripheral embolism recall rose from 49.53% to 63.55% (internal) and 57.73% to 72.16% (external).
• Patient-level AUC improvement was not statistically significant (0.947 vs 0.914, p=0.66); the pilot retrospective study needs prospective validation.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.