Comparative analysis of deep learning and radiomic signatures for overall survival prediction in recurrent high-grade glioma treated with immunotherapy.

Abstract

Background: Radiomic analysis of quantitative features extracted from segmented medical images can be used for predictive modeling of prognosis in brain tumor patients. Manual segmentation of the tumor components is time-consuming and poses significant reproducibility issues. We compare the prediction of overall survival (OS) in recurrent high-grade glioma(HGG) patients undergoing immunotherapy, using deep learning (DL) classification networks along with radiomic signatures derived from manual and convolutional neural networks (CNN) automated segmentation.

Materials and methods: We retrospectively retrieved 154 cases of recurrent HGG from multiple centers. Tumor segmentation was performed by expert radiologists and a convolutional neural network (CNN). From the segmented tumors, 2553 radiomic features were extracted for each case. A robust feature subset was selected using intraclass correlation coefficient analysis between manual and automated segmentations. The data was divided into a 9:1 ratio and validated through ten-fold cross-validation and tested on a rotating test set. Features selection was done by the Kruskal-Wallis test. The Radiomics-based OS predictions, generated using Support Vector Machine (SVM), were compared between the two segmentation approaches and against OS prediction by the CNN model adapted for classification. Model efficacy was evaluated using the area under the receiver operating characteristic curve (AUC).

Results: The clinical model AUC for OS prediction was 0.640 ± 0.013 (mean ± 95% confidence interval) in the training set and 0.610 ± 0.131 in the test set. The radiomics prediction of OS based on manual segmentation outperformed automatic segmentation (AUC of 0.662 ± 0.122 vs. 0.471 ± 0.086, respectively) in the test set. Robust features improved the performance of manual segmentation to AUC of 0.700 ± 0.102, of automated segmentation to 0.554 ± 0.085. The CNN prognosis model demonstrated promising results, with an average AUC of 0.755 ± 0.071 for training sets and 0.700 ± 0.101 for the test set.

Conclusion: Manual segmentation-derived radiomic features outperformed automated segmentation-derived features for predicting OS in recurrent high-grade glioma patients undergoing immunotherapy. The end-to-end CNN prognosis model performed similarly to radiomics modeling using manual-segmentation-derived features without the need for segmentation. The potential time-saving must be weighed against the lower interpretability of end-to-end black box modeling.