Abstract
Accurate prediction of IDH mutation status in gliomas is critical for guiding diagnosis, prognosis, and treatment planning. We enrolled 2,537 preoperative MRI of glioma patients (mean age 55.91 ± 14.79, 1,063 females) from 11 different datasets, consisting of 1,382 patients (mean age 58.26 ± 14.38, 548 females) in training set, 346 patients (mean age 57.43 ± 14.04, 141 females) in internal validation set, and 809 patients (mean age 53.92 ± 14.04, 374 females) in external test set, including 242 patients from The Cancer Genome Archive (TCGA) dataset. A fully automated Res3DNet model was established for isocitrate dehydrogenase (IDH) gene prediction. Four radiologists also read images from TCGA test dataset as a comparison with the deep learning model. Our Res3DNet model achieved AUCs of 0.946 (internal validation), 0.872 (external test), and 0.912 (TCGA test), with corresponding accuracies of 0.925, 0.806, and 0.840, respectively, outperforming ResNet model, I3D model, transformer model, and four radiologists.