Abstract
Aims
Treatment response assessment in glioblastoma is challenging. Patients routinely undergo conventional magnetic resonance imaging (MRI), but it has a low diagnostic accuracy for distinguishing between true progression (tPD) and pseudoprogression (psPD) in the early post-chemoradiotherapy time period due to similar imaging appearances. The aim of this study was to use artificial intelligence (AI) on imaging data, clinical characteristics and molecular information within machine learning models, to distinguish between and predict early tPD from psPD in patients with glioblastoma.
Method
The study involved retrospective analysis of patients with newly-diagnosed glioblastoma over a 3.5 year period (n=340), undergoing surgery and standard chemoradiotherapy treatment, with an increase in contrast-enhancing disease on the baseline MRI study 4-6 weeks post-chemoradiotherapy. Studies had contrast-enhanced T1-weighted imaging (CE-T1WI), T2-weighted imaging (T2WI) and apparent diffusion coefficient (ADC) sequences, acquired at 1.5 Tesla with 6-months follow-up to determine the reference standard outcome. 76 patients (mean age 55 years, range 18-76 years, 39% female, 46 tPD, 30 psPD) were included. Machine learning models utilised information from clinical characteristics (age, gender, resection extent, performance status), O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation status and 307 quantitative imaging features; extracted from baseline study CE-T1WI/ADC and T2WI sequences using semi-automatically segmented enhancing disease and perilesional oedema masks respectively. Feature selection was performed within bootstrapped cross-validated recursive feature elimination with a random forest algorithm and Naïve Bayes five-fold cross-validation to validate the final model.
Results
Treatment response assessment based on the standard-of-care reports by clinical neuroradiologists showed an accuracy of 33% (sensitivity/specificity 52%/3%) to distinguish between tPD and psPD from the early post-treatment MRI study at 4-6 weeks. Machine learning-based models based on clinical and molecular features alone demonstrated an AUC of 0.66 and models using radiomic features alone from the early post-treatment MRI demonstrated an AUC of 0.46-0.69 depending on the feature and mask subset. A combined clinico-radiomic model utilising top common features demonstrated an AUC of 0.80 and an accuracy of 74% (sensitivity/specificity 78%/67%). The features in the final model were age, MGMT promoter methylation status, two shape-based features from the enhancing disease mask (elongation and sphericity), three radiomic features from the enhancing disease mask on ADC (kurtosis, correlation, contrast) and one radiomic feature from the perilesional oedema mask on T2WI (dependence entropy).
Conclusion
Current standard-of-care glioblastoma treatment response assessment imaging has limitations. In this study, the use of AI through a machine learning-based approach incorporating clinical characteristics and MGMT promoter methylation status with quantitative radiomic features from standard MRI sequences at early 4-6 weeks post-treatment imaging showed the best model performance and a higher accuracy to distinguish between tPD and psPD for early prediction of glioblastoma treatment response.
Diagnostic performance of radiomics using machine learning algorithms to predict MGMT promoter methylation status in glioma patients: a meta-analysis
