Background: The importance of detecting genomic status of gliomas is increasingly recognized and IDH (isocitrate dehydrogenase) mutation and TERT (telomerase reverse transcriptase) promoter mutation have a significant impact on treatment decisions. Noninvasive prediction of these genomic status in gliomas is a challenging problem, however, deep learning model using magnetic resonance imaging (MRI) can be a solution. The image differences among facilities causing performance degradation, called domain shift, has been also reported in other tasks such as brain tumor segmentation. We investigated whether the gene status could be predicted by a deep learning model, and if so, to what extent it would be affected by domain shift.

Method: We used the data from the Multimodal Brain Tumor Segmentation Challenge (BraTS) and the Japanese cohort (JC) dataset consisted of brain tumor images collected from 544 patients in 10 facilities in Japan. We focused on IDH mutation and TERT promoter mutation. The deep learning models to predict the status of these genes were trained by the BraTS dataset or the training portion of JC dataset, and the accuracy of the models was evaluated by the test portion of JC dataset.

Results: The IDH mutation predicting model trained by the BraTS dataset showed 80.0% of accuracy with the validation portion of BraTS dataset, however, only 67.3% with the test portion of JC dataset. The TERT promoter mutation predicting model trained by the training portion of JC dataset showed only 49% of accuracy for the test portion of JC dataset.

Conclusion: IDH mutation can be predicted by deep learning models using MRI, but the performance degeneration by domain shift was significant. On the other hands, TERT promoter mutation could not be predicted enough by current deep learning techniques. In both mutations, further studies are needed.