Abstract
Recently identified changes in brain during daily MRI-guided radiation therapy (MRgRT) of patients with glioblastoma (GBM) can permit adaptive radiotherapy. However, daily changes can be subtle to detect and difficult to evaluate volumetrically without significant manual effort. The aim of this study is to develop a deep learning (DL) algorithm for automatic volume delineation of daily MRI scans from a hybrid MRI-RT system as an early warning system of tumor change and to facilitate analysis of delta radiomics and clinical trials based on MRI changes during RT.
GBM patients undergoing 30 fractions (6 weeks) of RT delivered by a combination MRI-RT system with concurrent temozolomide (TMZ) treatment were considered for study. The MRI data for each patient consisted of one predominantly T2-weighted balanced steady state free precession (bSSFP) planning scan (isotropic 1.5 mm3 resolution, 128 sec) followed by 30 daily bSSFP setup scans for 30 fractions of RT for each patient. Tumor lesion (TL) with/without surrounding edema, and resection cavity (RC) were manually contoured for each patient on each of the 31 scans in MIM. The DL-generated masks were converted to 3D volumes and compared using Dice Similarity Coefficient (DSC) between manual and automatic contours. In addition, the correlation between the automatic and manual volumes was examined with Person Correlation Coefficient (r).
The MRIs of 14 patients undergoing MRgRT with biopsy, subtotal resection, or gross total resection of GBM were incorporated into a dataset. Seven patients had only TL, 3 patients had both TL and RC, and 4 patients had RC only. MRIs from 11 patients were used in training and 3 patients were left for testing. Total of 8,505 axial slices were used to train the network, with 1,021 slices used as validation. DSC for TL and RC on the training and test datasets were: mean ± standard deviation: 0.87 ± 0.128 and 0.9 ± 0.122; 0.74 ± 0.233 and 0.8 ± 0.277, respectively. The CNN achieved DSC greater than 0.8 in 2 of the testing patients but less than 0.8 on 1, indicating that while successful, the network was not yet well generalized with the current training data. The automated lesion volumes correlated significantly with the manual volumes with r value of 0.9.
Contouring of brain lesions is a complex and time-consuming task, which impacts research throughput as it needs to be performed for every patient and scan. We have shown the promise of MASK R-CNN to produce automated contours for brain tumors. As more patients are treated with MRI-RT and with additional contrasts, the network will improve the accuracy and its generalizability. The use of DL in RT research improves the feasibility of visualizing tumor kinetics trends in real time to alert the physician to significant changes and further study of delta radiomics.