BACKGROUND
Glioblastoma is characterized by its extensive infiltration into the brain parenchyma, which poses challenges to accurate treatment volume targeting. Current treatment planning is based on the contrast-enhancing images. Tumor invasion, however, is well known beyond the contrast enhancement. Although deep learning shows potential in delineating tumor invasion based on multiparametric MRI, it requires large amount of training data. Image inpainting may reconstruct the lost image information based on image structure and requires less training data. The purpose of this study is to investigate whether a bi-level deep learning approach could integrate conventional neural network (CNN) and image inpainting to delineate tumor invasion based on the multiparametric MRI.
MATERIAL AND METHODS
A total of 115 patients (mean age 59.3 yrs, range 22 - 76 yrs, 87 males) with supratentorial de novo glioblastoma were recruited for maximal safe tumor resection. Post-operative temozolomide chemoradiotherapy was performed following the Stupp protocol. The Response Assessment in Neuro-oncology criteria was used to evaluate patient response. All MRI sequences were pre-operatively acquired using a 3T scanner, including T2-weighted, post-contrast T1-weighted, FLAIR, dynamic susceptibility contrast (DSC) and diffusion tensor imaging (DTI). The isotropic (p) and anisotropic (q) maps were generated from DTI. The relative cerebral blood volume (rCBV), mean transit time (MTT) and relative cerebral blood flow (rCBF) were generated from the DSC images after the leakage correction. All images were coregistered to post-contrast T1W images. Contrast-enhancing (CE) and non-enhancing (NE) regions were semi-automatically segmented and validated by an experienced neuro-radiologist. A tissue segmentation was performed using the FAST (FMRIB’s Automated Segmentation Tool) to segment the normal-appearing brain excluding the above both tumor regions and were classified into three tissue types (grey matter, white matter and CSF). A CNN model was applied to build a classification function between the CE region and the three normal-appearing tissue types based on the multi-parametric MRI. A bespoke inpainting scheme was then applied to predict the tumor invasion area in the NE region. Kaplan-Meier and Cox proportional regression were used to evaluate the significance of the predicted volume.
RESULTS
The volumes of the contrast-enhancing tumor is 53.6 ± 33.8 cm3. The predicted invasive volume in the non-enhancing region is 31.0 ± 17.7 cm3.The multivariate model shows that the predicted volume was significantly associated with OS (HR = 0.97, P= 0.006). The patient subgroup with a higher predicted volume showed significantly better OS (P= 0.009) and PFS (P= 0.021).
CONCLUSION
The proposed bi-level deep learning approach may effectively integrate multiparametric MRI for predicting tumor invasio.
