Thalamic volumetry via deep learning as an imaging biomarker in multiple sclerosis

The loss of thalamic volume, a central structure in the brain, is a crucial indicator of progressive nerve cell damage, often seen in conditions like Multiple Sclerosis (MS). While current clinical practice often utilizes T2-FLAIR MRI imaging to examine MS, existing methods are highly qualitative. This chapter aims to describe the development and validation of an innovative algorithm designed to measure thalamic volume using routine T2-FLAIR MRI scans. The proposed algorithmic biomarker is based on a 3D U-net using deep learning techniques. It was designed, trained, and validated 4590 MRI exams from 59 centers using an 80/10/10 train/validation/test split. A previously validated but not translationally applicable method called FIRST was used to generate thalamic masks from high-resolution 3D T1 images. The approach was then subjected to domain transfer to instead apply to clinical routine T2-FLAIR. Overall accuracy, interscanner reliability, and scan-rescan-reliability were assessed. Predictive value was also assessed via a longitudinal dataset with cognitive assessments. The algorithm was highly accurate, showing a 99.4% match with the results derived from 3D T1. Interscanner error and scan-rescan error were minimal. Crucially, the thalamic volume measured by the proposed approach was closely associated with disability and cognitive decline in patients, demonstrating significant predictive value. AI-based thalamic volumetry on clinical routine scans provides a robust, reliable, and relevant biomarker in multiple sclerosis.