To explore the application value of imaging omics based on non-contrast computed tomography (NCCT) in predicting extravasation of contrast agents during endovascular mechanical thrombectomy in patients with unilateral middle cerebral artery occlusion stroke.

A total of 96 patients with clinically diagnosed middle cerebral artery thrombosis and underwent endovascular mechanical thrombectomy in Suzhou kowloon Hospital, Shanghai Jiaotong University School of Medicine, were collected, including 67 patients in the exudative group and 29 patients in the non-exudative group. All patients were randomly assigned into a training set (67 cases) and a test set (29 cases) at a ratio of 7:3. The regions of interest of NCCT images were delineated using ITK-SNAP software to extract radiological characteristics. Mann-Whitney U test, Pearson correlation coefficient (Pearson), and the least absolute shrinkage and selection operator were used for feature selection. Models were trained and tested using Random Forest (RF) and Extreme Gradient Boosting (XGBoost). The ROC curves were plotted, and the area under the ROC curve (AUC), sensitivity, specificity, and optimal thresholds were calculated. The clinical utility of the models was further assessed through decision curve analysis (DCA).

Through screening, we obtained 6 imaging radiomics features that have the greatest predictive power. The training model of RF demonstrated the highest predictive efficiency. The AUC value of the model in the training group was 0.997, 95%CI: 0.990-1.000, and the accuracy, sensitivity, specificity and optimal thresholds were 97.0%, 97.8%, 95.2%, and 0.700, respectively. The DCA curve showed that the model threshold had a high clinical net benefit across a broad range, specifically from 0.5 to 0.9.

Machine learning based on NCCT demonstrates a high predictive value for the extravasation detection of contrast media during embolectomy of the middle cerebral artery.