Regression-Based Line Detection Network for Delineation of Largely Deformed Brain Midline

Abstract

Brain midline shift is often caused by various clinical conditions such as high intracranial pressure, which can be deadly. To facilitate clinical evaluation, automated methods have been proposed to classify whether midline shift is severe or not, e.g., larger than 5 mm away from the ideal midline. There are only limited methods using landmark or symmetry, attempting to provide more intuitive results such as midline delineation. However, landmark- or symmetry-based methods could be easily affected by anatomical variability and large brain deformations. In this study, we formulated the midline delineation as a skeleton extraction task and proposed a novel regression-based line detection network (RLDN) for the robust midline delineation especially in largely deformed brains. Basically, the proposed method includes three parts: (1) multi-scale line detection, (2) weighted line integration, and (3) regression-based refinement. The first two parts were used to capture high-level semantic and low-level detailed information to extract deformed midline, while the last part was utilized to regress more accurate midline positions. We validated the RLDN on 100 training and 28 testing subjects with a mean midline shift of 7 mm and the maximum shift of 16 mm (induced by hemorrhage). Experimental results show that our proposed method achieves state-of-the-art accuracy with a mean line difference of mm and F1-score of 0.78 from manual delineations. Our proposed robust midline delineation method is also beneficial for other cases such as midline deformation from tumor, traumatic brain injury, and abscess.