BRMP-04. AI-based biomarker of the peritumoral region using tissue microstructure

Abstract

Abstract PURPOSE Glioblastomas, the most common malignant brain tumor [BS1], infiltrate into peritumoral brain structures, making clinical management challenging. An unmet need is to develop a biomarker that reliably characterize infiltration in the peritumoral region, where surgical biopsy or resection may be hazardous. Diffusion tensor imaging (DTI) with multicompartment modeling can characterize extracellular free water, providing unique information of the tissue microstructure that is able to capture this heterogeneity. We propose a novel biomarker based on peritumoral tissue microstructure, using deep-learning on DTI-based free water map. METHOD Peritumoral regions were automatically segmented for 136 patients with brain tumors (86 glioblastomas and 50 metastases, ages 23–87 years, 65 females). We trained a Convolutional Neural Network (CNN) on free-water maps using automatically defined patches in the peritumoral area from glioblastomas and metastases, labeled as low free-water and high free-water to extract a microstructural index for each voxel. To extract the biomarker, we grouped peritumoral voxels into connected components (CCs) where adjacent voxels have high (>0.9) microstructural index values. Two independent test sets related to two clinically significant problems were evaluated: i) metastases vs. glioblastomas; ii) glioma patients categorized into short and long survival groups and the number of CCs were statistically compared. RESULT Two-sample t-tests showed significant group difference in the number of CCs between metastases and glioblastomas (p< 0.05), and short and long-survivals (p<0.05) with higher number of CCs in metastases and long-survivals, which suggests smaller number of voxels in CCs. CONCLUSION The proposed biomarker based on CCs of microstructural index captures the differences in infiltration of the peritumoral region, showing larger CCs in glioblastomas and short-survivals corresponding to higher infiltration. CLINICAL IMPORTANCE The proposed biomarker provides a novel insight into the peritumoral microenvironment and can be derived from clinically feasible DTI data, providing new possibilities for the diagnosis and treatment of glioblastoma.