Health & body · 2017
Can AI predict a 's from conventional MRI scans?
Yes—in this multi-institutional study, a deep learning model predicted IDH status with 85.7% accuracy in the testing set (AUC = 0.94). Adding age at diagnosis increased testing accuracy to 89.1% (AUC = 0.95).
Details
Isocitrate dehydrogenase (IDH) mutations in patients with glioma are associated with longer survival and may help guide treatment decisions. The researchers aimed to predict glioma IDH status from preoperative MRI using a residual convolutional neural network.
This multi-institutional prediction-model development study used preoperative imaging from 201 patients at the Hospital of University of Pennsylvania (HUP), 157 patients at Brigham and Women's Hospital (BWH), and 138 patients from The Cancer Imaging Archive (TCIA). The data were divided into training, validation, and testing sets.
The researchers trained a residual convolutional neural network for each MRI sequence—FLAIR, T2, T1 before contrast, and T1 after contrast—and built a predictive model from the networks' outputs. To enlarge the training set and prevent overfitting, they augmented the training images with random rotations, translations, flips, shearing, and zooming.
The neural network model achieved IDH prediction accuracies of 82.8% in the training set (AUC = 0.90), 83.0% in the validation set (AUC = 0.93), and 85.7% in the testing set (AUC = 0.94). When age at diagnosis was added to the model, accuracy increased to 87.3% in the training set (AUC = 0.93), 87.6% in the validation set (AUC = 0.95), and 89.1% in the testing set (AUC = 0.95).
The researchers concluded that they had developed a deep learning technique to noninvasively predict IDH genotype in grade II–IV glioma using conventional MRI and a multi-institutional data set.
Publication information: Clin Cancer Res; 24(5); 1073–81. ©2017 AACR.
What do you think?
Good conversations begin with curiosity.
You can read comments without an account. Sign in to write.
Loading comments…