NeuroAnatomy-Aware Refinement in Intracerebral Hemorrhage Segmentation and Towards Traumatic Brain Injury Severity Assessment

Abstract

Timely and objective traumatic brain injury (TBI) severity assessment is essential for clinical decision-making and for integrating imaging into digital health workflows. This paper presents a multimodal framework that combines quantitative biomarkers derived from non-contrast head CT with structured clinical variables to predict TBI severity. On the imaging branch, a slice-based segmentation approach is adopted to handle heterogeneous CT volumes with varying numbers of axial slices, together with an anatomy-aware refinement step to improve label consistency and anatomical plausibility of hemorrhage recognition. The extracted imaging biomarkers are then integrated with key clinical indicators and used for severity stratification by conventional machine-learning classifiers. Experiments on a matched multimodal dataset consisting of head CT images and structured clinical and tabular data demonstrate that incorporating segmentation-derived imaging features improves the prognostic assessment of TBI severity compared with using clinical variables alone. These findings highlight the added value of lesion quantification from CT for multimodal severity prediction and support the practicality of the proposed framework for clinical severity assessment.