Automated Deep Learning-Based Image Segmentation for the Correction of Normal Brain Dose-Volume Parameters in Brain SRS for Cerebral AVMs

Abstract

Brain radiation necrosis (RN) is a dose-limiting toxicity of stereotactic radiosurgery (SRS) for cerebral arteriovenous malformations (AVMs). Prior studies show a significant relationship between the probability of RN and the volume of brain receiving high doses, in particular, the volume receiving at least 12Gy (V12). However, due to the prohibitive time cost of manually segmenting brain from other intracranial tissues, prior studies have treated all intracranial tissues homogenously as brain. Large AVMs can cause significant anatomical distortions, with dilated ventricles, draining veins, and embolized nidus elements potentially displacing brain tissue from the high dose volume. We hypothesized that there would be a significant difference between estimated V12 for all intracranial tissues (V12ic) and V12 for brain parenchyma specifically (V12brain) in AVM cases. To assess the difference between V12ic and V12brain (ΔV12), we used a convolutional neural network (CNN) to auto-segment magnetic resonance images (MRIs) from SRS-treated patients into vascular structures, CSF, and brain tissue, fusing the segmented images with treatment plans to calculate V12ic and V12brain. SRS plans from 8 AVM patients who underwent single stage SRS and 4 who underwent multistage SRS were included. For multistage SRS plans, each stage was treated as an independent plan, for a total of 16 analyzed treatment plans. A previously trained CNN was used to automatically segment multiparametric pre-treatment MRIs into arteries, veins, CSF, embolized vessels, and brain parenchyma. Segmentation labels were converted into contours and registered to treatment plans. Dose-volume histograms were generated, from which V12ic and V12brain for all intracranial tissue and brain parenchyma, respectively, were analyzed. We assessed the relationship between ΔV12 (V12ic - V12brain) and target volume using linear regression. Median target volume was 1.75cc (range [0.22-9.95cc]). Median V12ic was 8.31cc (range [2.1-21.4cc]) while median V12brain was 4.0cc (range [1.3-13.3cc]). Median ΔV12 was 3.7cc (range [0.46-9.4cc]). There was a significant linear relationship between target volume and ΔV12 (Pearson r2 = 0.57, p < 0.01), with larger targets associated with a greater ΔV12. Anatomical distortions produced by an AVM nidus can produce a significant difference between V12ic and V12brain, an effect that is more pronounced for larger lesions. Further studies are needed to directly assess the relationship between V12brain and likelihood of post-SRS RN; however, estimating dose to brain tissue specifically may help identify patients who can safely receive dose-escalated therapy as well as patients who require a more conservative approach. CNN-based automated segmentation provides a means of overcoming the previously prohibitive time cost of performing such an analysis.