MSU Researchers Teach AI to Detect CT Scan Abnormalities Without Manual Annotation

Researchers at the Artificial Intelligence Center of Lomonosov Moscow State University (MSU) have developed a self-supervised method that enables a neural network to detect abnormalities in computed tomography (CT) scans without requiring manually annotated training images.

The new approach could significantly accelerate AI model development. Conventional medical imaging systems are typically trained on manually labeled scans, but such datasets are often limited and incomplete. In many cases, scans are annotated for one disease while other abnormalities remain unlabeled, making it difficult to build AI systems capable of recognizing a broad range of pathologies.

The method was evaluated on four large medical imaging datasets containing CT scans with signs of lung cancer, pneumonia, and liver and kidney tumors. It demonstrated substantially better performance than existing approaches.