Harnessing Radiomics: Enhancing Prognosis in Pulmonary Embolism via Thrombus Imaging

A new primary study identifies CT-derived thrombus radiomics as a robust prognostic tool, demonstrating that thrombus imaging materially improves risk classification in pulmonary embolism.

Radiomics-derived thrombus signatures outperformed conventional CT markers for 30‑day mortality discrimination in the analyzed cohort.

In a retrospective cohort of 86 CT pulmonary angiograms, investigators used 30‑day mortality and troponin level as primary endpoints and performed manual segmentation of each arterial thrombus. They extracted 3D Shape and first‑order measures plus GLCM texture metrics and identified Image‑Mean, firstorder_InterquartileRange, firstorder_10Percentile and glcm_ClusterShade as the most predictive features in this dataset. These thrombus shape and texture associations produced high cohort‑level discrimination (3D mortality AUC ≈ 0.969; overall AUCs ≈ 0.9686) and represent demonstrated within‑sample associations rather than external validation.

The investigators' pipeline combined correlation filtering with 12 feature‑selection methods reduced to a 10‑feature set and used 5‑fold cross‑validation to limit within‑sample overfitting. The study authors' analysis evaluated 12 classifiers (including SVM, XGBoost and Ensemble Bagged Trees) and reported top performances from SVM, XGBoost and ensemble bagged trees, with in‑cohort AUCs and accuracies up to ~0.97.

Key operational barriers include the manual segmentation workload, CT acquisition heterogeneity (contrast timing, slice thickness and reconstruction kernels), and the absence of prospective multi‑center external validation. Because protocol variability can alter feature stability, mitigation steps include standardized reconstruction parameters, automated segmentation with QC flags, and prospective multi‑site benchmarking. As applied today, radiomics outputs are best introduced as adjunct risk scores integrated into PACS/reporting pipelines rather than as standalone decision engines; regulatory, IT and workflow alignment will be necessary but are tractable for pilot deployments.