Automatic classification of Traumatic Brain Injury in Emergency CT Imaging

Overview & Synopsis

Traumatic brain injury impacts (TBI) around 150,000 persons each year in Europe, leading to 8,000 deaths and causing health complications akin to those of a chronic disease. Computed tomography imaging is used to evaluate head trauma, as it is the most efficient and quickest method for patients requiring heavy treatment. The diagnosis of this medical condition must be performed quickly to avoid aggravated long-term repercussions. This pathology remains challenging to diagnose using CT imaging because of the low contrast resolution of soft tissues. It may also be mistaken for other conditions, such as a hemorrhagic stroke. It is therefore essential to quickly assess all incoming patients and guide them to the appropriate department. To this aim, this thesis examines how the use of artificial intelligence (AI) can help alleviate the difficulties of low contrast CT scans. A technique for extracting more information from these images involves the use of radiomic metrics, which enables the quantitative characterization of qualitative data. As a first step, this thesis investigates whether the intrinsic variability of the extracted radiomics features needs to be corrected before any analysis. It then focuses on the development and evaluation of segmentation methods, as well as machine-learning classifiers on an all-comers population to differentiate traumatic hemorrhage from medical hemorrhage, ischemia, or healthy scans. Finally, the thesis explores the possibility of using radiomic metrics to automatically classify the severity of injuries according to the Abbreviated Injury Scale (AIS).