Abstract
Distinguishing between true progression (TP) and pseudoprogression (PP) post-radiotherapy (RT) is of paramount importance for treatment management of patients with glioblastoma (GBM). MR-Linac systems allow for daily monitoring of tumor changes throughout the course of RT. We hypothesized that the patterns of tumor volume change during RT may enable early prediction of treatment response.
Using an IRB-approved prospective cohort of GBM patients undergoing 30 fractions of chemoRT to 60 Gy on a 0.35T MR-Linac, tumor/edema (tumor lesion) regions of interest (ROI) were contoured on daily T2-weighted treatment set-up scans. The obtained tumor lesion (TL) volume changes during treatment were smoothed with a moving average Gaussian window over time. Non-negative Matrix Factorization (NMF) was applied to the data matrix D (N x F), containing the trajectories in its rows for each patient, where N is the number of patients analyzed and F is the number of fractions. NMF represents D as a linear combination of three temporal (hidden) patterns and their weights in each individual trajectory. The same analysis was performed for ΔD, containing the changes in volumes with reference to the first fraction. The calculated weights were scaled in [0, 1], expressed as probabilities (by ℓ1-normalization) and used as features in Linear Discriminant Analysis (LDA). The LDA model was trained to differentiate between no progression (NP), PP and TP, and assessed by leave-one-subject-out cross-validation.
Thirty-six patients were screened for inclusion: 9 were excluded due to no T2 lesion (resection cavity only). Of the remaining 27 GBM patients analyzed, 10 had no tumor growth on first post-RT diagnostic MRI, 6 were determined to have PP based on regression or long-term stability of findings, and 11 had TP due to continued progression of disease past 6 months, rapid patient death from disease, or tissue sampling showing active malignancy. With the use of only 2 features, LDA achieved an overall accuracy of 70.4% classifying correctly: 6 (60%), 4 (67%), and 9 (82%) patients with NP, PP, and TP, respectively. The temporal NMF patterns (monotonous decrease, rapid increase during the third part of the treatment, etc.) indicate that there is enough signal to classify patients’ response based on the pattern tumor volume changes during RT.
We identified tumor dynamics’ patterns during RT, indicative of differential behavior of tumor growth between TP and PP. Although with a limited number of patients, these initial results suggest that tumor volume changes during treatment may provide early markers of treatment response. This could allow physicians to adapt/intensify treatment in real time for poorly responding patients. Next steps include automating the process of real-time tumor volume monitoring by incorporating a deep learning solution for automatic volume delineation on daily treatment set-up scans.