An Artificial Intelligence-Based Screening Tool for Detection of Cerebral Venous Thrombosis on Non-Contrast Brain CT

While non-contrast computed tomography (NCCT) is the primary imaging modality in emergency settings, it has a low sensitivity for detection of cerebral venous thrombosis (CVT). This study aimed to develop an artificial intelligence (AI) tool to screen for CVT on routine NCCT scans.

We used a retrospectively collected dataset of 692 patients, including 258 with CVT from Iran Cerebral Venous Thrombosis Registry (ICVTR) (code: 9001013381) and 434 controls. NCCT images were processed using the pretrained CT foundation model to generate high-dimensional embeddings. These embeddings were used to train several machine learning classifiers: support vector machine (SVM) with linear, polynomial, and radial basis function kernels, logistic regression, random forest, an artificial neural network, and a soft voting ensemble model. Model performance was assessed using a 10-fold cross-validation protocol.

The random forest model achieved the highest specificity of 0.856 ± 0.050, excelling at identifying non-CVT cases. Conversely, the ensemble model yielded the highest sensitivity of 0.614 ± 0.087. The SVM with a linear kernel provided the best overall discriminative ability, with the highest area under the receiver operating characteristic curve of 0.718 ± 0.053. No single model demonstrated superiority across all metrics, reflecting the inherent challenge of detecting CVT on NCCT.

Our findings demonstrate the feasibility of using an AI-based model to detect CVT on non-contrast CT scans. While not yet a replacement for expert radiological interpretation, this approach serves as a promising automated screening tool. It has the potential to reduce diagnostic delays and improve patient outcomes by flagging suspicious cases in emergency workflows.