Brain midline shift detection and quantification by a cascaded deep network pipeline on non-contrast computed tomography scans

Brain midline shift (MLS), demonstrated by imaging, is a qualitative and quantitative radiological feature which measures the extent of lateral shift of brain midline structures in response to mass effect caused by hematomas, tumors, abscesses or any other space occupying intracranial lesions. It can be used, with other parameters, to determine the urgency of neurosurgical interventions and to predict clinical outcome in patients with space occupying lesions. However, precisely detecting and quantifying MLS can be challenging due to the great variability in clinically relevant brain structures across cases. In this study, we investigated a cascaded network pipeline consisting of case-level MLS detection and initial localization and refinement of brain landmark locations by using classification and segmentation network architectures. We used a 3D U-Net for initial localization and subsequently a 2D U-Net to estimate exact landmark points at finer resolution. In the refinement step, we fused the prediction from multiple slices to calculate the final location for each landmark. We trained these two U-Nets with the Gaussian heatmap targets generated from the brain's anatomical markers. The case-level ground-truth labels and landmark annotation were generated by multiple trained annotators and reviewed by radiology technologists and radiologists. Our proposed pipeline achieved the case-level MLS detection performance of 95.3% in AUC using a testing dataset from 2,545 head non-contrast computed tomography cases and quantify MLS with a mean absolute error of 1.20 mm on 228 MLS positive cases.