A deep learning based radiomics model for differentiating intraparenchymal hematoma induced by cerebral venous thrombosis

This research seeks to formulate and confirm a deep learning radiomics nomogram (DLRN) based on nonenhanced CT (NECT) to differentiate intraparenchymal hematomas associated with Cerebral Venous Thrombosis (CVT) from those caused by other etiologies. 275 patients with intraparenchymal hematomas who underwent NECT were included in this work. Participants from two medical centers were assigned to distinct cohorts: a training set from Center 1 consisting of 192 patients (46 with confirmed CVT and 146 with other etiologies) and an external test set from Center 2 comprising 83 patients (24 with confirmed CVT and 59 with other etiologies). Conventional radiomics (Rad) features and deep learning (DL) features were derived from NECT images and integrated to form deep learning radiomics (DLR) features. Separate predictive models were constructed using Rad, DL, and DLR features. A DLR signature was obtained and integrated with medical characteristic variables to develop the DLRN model via multivariate logistic regression. The model's predictive performance was evaluated using ROC curves and decision curve analysis (DCA). Sixteen Rad features and three DL features were selected to construct the fused DLR features. The DLR model exhibited superior discriminative performance in identifying secondary intraparenchymal hemorrhage due to CVT compared to the individual Rad and DL models, achieving an AUC of 0.904(95% CI: 0.8207-0.9879) in the external test cohort. By integrating the DLR signature with epilepsy, the DLRN model was developed, demonstrating the highest predictive accuracy among all radiomics models, with an AUC of 0.911(95% CI: 0.8265-0.9962) in the external test cohort. Decision curve analysis (DCA) revealed that the DLRN model offered enhanced practical applicability relative to other radiomics-based models. A CT-based DLRN model was developed to distinguish intraparenchymal hematomas associated with CVT and those caused by other factors. The model offers a rapid and non-invasive diagnostic approach without the need for contrast enhancement, potentially improving early diagnosis and clinical decision-making.