Comparative study of radiologists vs machine learning in differentiating biopsy-proven pseudoprogression and true progression in diffuse gliomas

Abstract

Background and Purpose: MRI features of tumor progression and pseudoprogression may be indistinguishable especially without enhancing portion of the diffuse gliomas. Our aim is to discriminate these two conditions using radiomics and machine learning algorithm and to compare them with human observations. Materials and Methods: Three consecutive MRI studies before a definitive biopsy in 43 diffuse glioma patients (7 pseudoprogression and 36 true progression cases) who underwent treatment were evaluated. Two neuroradiologists reviewed pre- and post-contrast T1, T2, FLAIR, ADC, rCBV, rCBF, K2, and MTT maps. Patterns of enhancement, ADC maps, rCBV, rCBF, MTT, K2 values, and perilesional FLAIR signal intensity changes were recorded. Odds ratios (OR) for each descriptor, raters' success in predicting true and pseudoprogression, and inter-observer reliability were calculated using the R statistics software. Unpaired Student's t-test and receiver operating characteristic (ROC) analysis were applied to compare the texture parameters and histogram analysis of pseudo- and true progression groups. All first-order and second-order image texture features and shape features were used to train and test the Random Forest classifier (RFC). Observers' success and RFC were compared. Results: Observers could not identify true progression in the first visit. However, accuracy of the RFC model was 81%. For the second and third visits, the rater's success of prediction was between 62% and 72%. The accuracy for the second and last visit with RFC was 75% and 81%. Conclusions: Random Forest classifier was more successful than human observations in predicting pseudoprogression using MRI.

Keywords

• Pseudoprogression
• Tumor progression
• Machine learning
• Radiomics
• Glial tumor