Radiomics and machine learning in the differentiation of pseudoprogression and true tumor progression in high-grade gliomas

Purpose: Patients undergoing treatment for high-grade gliomas may exhibit indistinguishable MRI ndings of true tumor progression from pseudoprogression. We employed to shape and texture features on a machine learning algorithm to make accurate predictions compared to human observations. Materials and Methods: Three consecutive studies prior to a denitive biopsy in 50 high-grade glioma patients who underwent treatment were evaluated. The last study before revision surgery with a known pathology result (12 pseudoprogression and 38 true progressions) was chosen as the nal time point. Pre-post-contrast T1, T2, FLAIR, ADC, CBV, CBF, K2, MTT maps were reviewed by 2 neuroradiologists. Patterns of enhancement, ADC maps, CBV, CBF, MTT, K2 values, perifocal FLAIR signal intensity changes were recorded. Odds ratios (OR) for each statement, raters' success of predicting true and pseudoprogression and inter-observer reliability were calculated with R statistics software. To compare the texture parameters and histogram parameters of pseudo and true progression groups, the unpaired Student's t-test and receiver operating characteristic (ROC) analysis were applied. All rst-order and second-order image texture features and shape features were used for training and testing Random Forest classier (RFC). Results: For the first visit, none of the observers could identify true progression. For the first and second visits, the rater's success of prediction for the visits before biopsy were %72 and %62. The AUC for the same visits with RFC was %81, the success by using first order texture analysis alone was between %61 and %68. For the first visit, the AUC of the RFC model was %70 where second-order texture analysis could distinguish the groups between %65 and %67. Conclusions: Radiologists success by using both conventional MRI and perfusion maps was lower than RF classier success using conventional MRI images alone. Second order texture features were useful to make better predictions than human raters.