Return To Abstract Listing

Scientific Session 29 — Efficacy/Administration/Informatics - Appropriateness

Friday, May 5, 2017

Abstracts 2441-3140



2907. Brain MRI Sequence Identification Using Deep Learning

Garg N1*,  Wei P2,  ElShikh M1,  Rayan J3,  Li X3 1. University of Texas MD Anderson Cancer Center, Houston, TX; 2. University of Texas Health Science Center, Houston, TX; 3. Baylor College of Medicine, Houston, TX

Address correspondence to N. Garg (ngarg@mdanderson.org)

Objective: Deep learning is a hot topic of research in computer vision and artificial intelligence that uses neural networks combined with massive computing power. Our goal was to see if deep learning can be used to identify T1- and T2-weighted MRI sequences in brain images.

Materials and Methods: We selected 100 T1- and 100-weighted T2 axial DICOM images of the brain from the glioblastoma multiforme dataset of The Cancer Genome Atlas by the National Cancer Institute. We then sampled 15,000 random 100 x 100 pixel patches from these images. We split the samples into 8000 pixel patches for training and 7000 for testing. We trained on three different neural network topologies: softmax regression (fully connected; 10,000, 2 neurons), two-layer support vector machine (SVM) (10,000, 400, 50, 2 neurons), and a deep learning network with two restricted Boltzmann machine (RBM) layers (400 and 200 neurons) followed by two SVM layers (100 and 50 neurons) in between the input (10,000 neurons) and output (2 neurons) layers. All experiments were done on a PC with the Nvidia Titan X graphics processing unit and CUDA 7.5 software (Nvidia Development). Code was written in Python for preprocessing and C++ for training compiled with Microsoft Visual Studio 2013 via ArrayFire. All T1- and T2-weighted images as well as sampled patches were reviewed by a radiologist to make sure the data were clean and accurate.

Results: We achieved an accuracy of 80% on test data and 99% on training data using the deep learning network after training for 2 minutes. We achieved a similar accuracy with a two-layer SVM, but it took 2 hours of training. A single softmax regression layer network was not effective, giving only 55% accuracy on test data and 63% accuracy on training data after converging the error rate and 10 minutes of training.

Conclusion: Sequence classification is one of the first things a radiologist confirms when looking at MR images. We successfully built and trained a reasonably accurate deep learning classifier for T1- and T2-weighted brain MRI sequences with a fairly small amount of data in a short time. We are encouraged by these results without any feature engineering or hand-crafted algorithms, and the strength of deep learning approaches. We gained insight into tuning the learning rate and other hyperparameters as well as network topologies useful for medical image work.