FIRST LOOK: Using Residual Convolutional Neural Networks to Improve Treatment of Brain Tumors
Gliomas are the most common primary brain tumor in adults and the cure remains elusive. Certain pathological biomarkers including the mutation status of isocitrate dehydrogenase (IDH) have significant implications on the benefits of aggressive surgical resection, response to treatment and overall prognosis. Tissue sampling via biopsy or surgical resection has been required to elucidate biomarker status.
Our research is designed to predict IDH status of gliomas using preoperative structural MR imaging by utilizing a residual convolutional neural network.
Preoperative imaging for patients with grade II-IV gliomas was obtained from multiple institutions including Brigham and Women’s Hospital, the Hospital of the University of Pennsylvania and The Cancer Imaging Archive. The dataset was divided into training, testing and validation sets. A residual convolutional neural network was trained from each MR sequence to build a predictive model. Data augmentation in the form of image rotation, translations, flips, shearing and zooming was utilized. The age at the time of diagnosis was also incorporated into the model.
The trained neural network achieved IDH prediction accuracy of 87.3% (AUC 0.93), 87.6% (AUC 0.95), and 89.1% (AUC 0.95) with training, validation and training sets respectively.
A multi-institutional dataset was used to generate and validate a deep learning model capable of predicting with high accuracy IDH mutation status in patients with low and high grade gliomas. As a result, we now have a tool that can be deployed in the clinical setting allowing us to better counsel patients and tailor the surgical approach and goal to the individual. Furthermore, this project serves as a proof-of-concept for a suite of tools that will use deep learning to predict various biomarkers in patients with brain tumors, augmenting traditional radiological interpretation for clinical decision making.
For more information about Dr. Arnaout’s research, please contact Partners HealthCare Innovation by clicking here.
Figure 1. A, Image preprocessing steps in our proposed approach. B, A modified 34-layer residual neural network architecture was used to predict IDH status. C, The learning rate schedule. The learning rate was set to 0.0001 and stepped down to 0.25 of its value when there was no improvement in the validation loss for 20 consecutive epochs.
Figure 2. ROC curves for training, validation, and testing sets from training on three patient cohorts for age only (A), combining sequence networks (B), and combining sequence networks + age (C). The testing set AUC for combining sequence networks + age was 0.95.
Most Recent Posts:
The Medically Engineered Solutions in Healthcare (MESH) Incubator: A case study of innovation “push”
In an opinion piece for STAT, Marc Succi, MD, Clinician-in-Residence at Mass General Brigham Innovation, addresses the…