Automatically segment the vestibular schwannomafrom pre- and post-GKRS MR images using dual-pathway model, and calculate the MR radiomics based on the ROIs obtained from model. According to the extracted features, we could predict the treatment response for the GKRS.We retrospectively collected the MR images and tumor contours of 516 participants with VS, these had been obtained by the Gamma Knife team at Taipei Veterans General Hospital, Taiwan. The trained dual-pathway U-Net model used multi-parametric MR images (T1-weighted, T1-weighted with contrast, and T2-weighted) as training input, achieving the testing dice score: 0.90.
We calculated the MR radiomics from pre-GKRS images to predict the treatment response and whether the pseudo-progression will occur. After the statistical analysis and the feature selection to retain the most important features, the two-level machine-learning models were built by these features. The two-level classification model respectively achieved the testing AUC: 0.877 and 0.816 for each level.The assessment system includes the image pre-processing workflow, automatic segmentation and treatment response prediction algorithm. This pipeline is not only useful for clinical physicians to assess the VS treatment response, and also can be adopted to other kinds of benign and malignant brain tumor to build up their own assessment systems. This technique provides another health care solution that can be implemented at the devices of radiotherapy/radiosurgery to assist the treatment planning.