Abstract TMP49: Radiomics Model Predicts Histological Thrombus Composition on Admission CT Imaging

Abstract

Background: Red blood cells (RBC)-rich thrombus are more easily retrieved via endovascular procedures while platelet-rich thrombus are more resistant to recanalization. Our aim was to generate a radiomics model able to identify both RBC and platelet-rich thrombus at CT admission in patients undergoing mechanical thrombectomy. Methods: We included consecutive patients that received mechanical thrombectomy due to a large vessel occlusion in which thrombi was obtained. Thrombi obtained during the procedure were hematoxiline-eosine processed and proportions of RBC were determined. Relative proportion of the platelets in the thrombi was quantified by using a immunohistochemical staining recognizing CD61. We considered RBC-rich thrombi those with a content of RBC>30% and platelet-rich thrombi those with a content of CD61>70%. Thrombi were segmented manually on co-registered non-contrast CT (NCCT) and CT angiography (CTA) at admission (<1mm thickness). 804 quantitative radiological features were extracted from the NCCT and CTA images. The most relevant features were selected with a genetic algorithm. Xgboost models were trained and one-versus-all cross-validated on the selected features to classify RBC and platelet-rich thrombi. Results: From 97 patients with histologically and radiologically available thrombi, 43% were female and mean age was 71(±12) years. The area under the curve for detecting RBC>30% was 0.938 (sensitivity 86%, specificity 86%, PPV 84%). Regarding platelet-rich thrombus, the area under the curve for detecting CD61>70% was 0.89 (sensitivity 83%, specificity 84%, PPV 74%). Feature importance by imaging type for RBC-rich and platelet rich thrombus was 70.33% and 69.8% for NCCT and 29.67% and 30% for CTA, respectively. The most important feature types for RBC-rich thrombus were histograms (51%) and first order (27%) while the most important features for platelet rich-thrombus were histograms (94%) followed by texture (5.94%). No information about shape improved any model. Conclusion: Our radiomics model can reliably identify RBC and platelet-rich thrombi. Fast identification of thrombus histological components on CT at arrival can help to design the preferred therapeutic strategy