An Automated MRI Analysis Tool to Measure the Tumor Volume and Assess the Treatment Response for Glioblastoma

Access full-text files

Date

2022-09-29

Authors

Kabir, Tanjida
Hsieh, Kang-Lin
Nunez-Rubiano, Luis
Cai, Yu
Hsu, Yu-Chun
Quintero, Juan C. Rodriguez
Arevalo, Octavio
Zhao, Kangyi
Zhang, Jackie Jiaqi
Zhu, Jay-Jiguang

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

Glioblastoma is the most common and aggressive grade IV glioma tumor. During surgical resection, complete tumor removal is impossible due to irregular structure and can be infiltrative into the adjacent brain tissue. Therefore, measuring residual tumor volumes and their detailed location becomes a vital predictor of patients’ survival. Furthermore, radiologists manually segment tumor regions from normal brain tissue and compare the pre-surgery and follow-up MRIs to conduct post-surgery evaluations. This process is time-consuming and operator-dependent due to postoperative changes and tumors' irregularity. Therefore, the overarching aim of this work is to design an artificial intelligence (AI) framework to estimate the residual tumor volume considering the brain structural variations and assess the efficacy of the therapy utilizing imaging and clinical features. We used 419 pre-surgical and 310 follow-up segmented MRIs. All cases include four MRI modalities- T1, T1+Gd, T2, T2-FLAIR. Our study demonstrates that because of the morphological changes in the post-surgical brain, the state-of-the-art AI segmentation models trained on pre-surgery MRIs suffer from a 3-20% performance drop on follow-up MRIs. Besides, these models show a significant drop in generalizability and consistency on independent MRIs (p-value<0.05), indicating the necessity of training follow-up MRI-based segmentation models to assess the treatment response. We propose an encoder-decoder-based segmentation model where encoders utilize contrastive learning schemes to identify tumor location and shape by consolidating domain-specific and problem-specific features to reduce variation of the tumor’s irregularity. We replace the decoder’s deterministic layers with Bayesian layers to quantify the uncertainty in the model's prediction. The proposed model achieved >0.85 dice score on average for segmenting and measuring the volume of different tumor sub-regions- edema, enhancing, and non-enhancing regions. These areas’ prompt and accurate identification is crucial to measure residual disease, detect tumor recurrence, and identify treatment-associated side effects.

Description

LCSH Subject Headings

Citation