Multi-Sequence MRI radiomics model for discrimination of recurrence and pseudoprogression in gliomas

The clinical treatment strategies for glioma recurrence and pseudoprogression are completely different. Precise differentiation between these conditions is essential and radiomics serves as a mature solution to bridge this gap.

We retrospectively analyzed MRI data from 240 diffuse glioma -- 112 from our Center1 and 59 from the UPENN dataset for training, plus an independent test cohort of 69. Radiomics features from T1WI, T2WI, T1CE, FLAIR, DWI(b = 1000), and ADC maps were extracted and used in feature engineering and machine learning to differentiate glioma from pseudoprogression. In the training set, the prediction performance was evaluated by using accuracy and the area under the curve (AUC) of the receiver operating characteristics, while the model stability was analyzed by using the relative standard deviation of the AUC (RSD_AUC).

We extracted features by combining six sequences, and after feature engineering, 45 features remained. We then applied these features to ten machine learning models and finally validated them in a separate training set. Support Vector Machine (SVM) achieved favorable prediction performance (AUC: 0.779) whereas Multilayer Perceptron (MLP) displayed high performance (AUC: 0.88 ± 0.05, accuracy: 0.81 ± 0.04) and high robustness against data perturbation (RSD_AUC: 5.25 %) CONCLUSIONS: Imaging models effectively differentiate between glioma recurrence and pseudoprogression, warranting their utilization in clinical practice.