Andrew M. Sohn
School:
Sidney Kimmel Medical College at Thomas Jefferson University
NIH Institute:
NIH Clinical Center (CC)
Research Project Title:
Deep Organs: Toward Automated Abdominal Organ Segmentation Using Deep Learning
Research Summary:
Using new and highly successful techniques in deep learning, convolutional neural networks (CNN), and other computer vision techniques, we sought to classify, localize, and segment abdominal organs automatically. A strong motivation in using deep learning was to avoid using image registration-based techniques (i.e. ATLAS) for classification and segmentation, which consists of high computational costs (high-performance computing), hand-crafted features, and limited flexibility of what kind of dataset can be analyzed.
Using a dataset from the Beyond the Cranial-Vault Synapse Challenge, with IRB approval, 50 labeled CT scans from two clinical studies weres used to train the CNN. Upon extracting the features from the CNN, additional features from a landmark detection algorithm were incorporated. This allows for a more structured input (spatial locations of the organs in the abdomen) to obtain more precise feature extraction and further minimize over-fitting. Finally, classification was done using a multi-class linear support vector machine (SVM). Software development was performed on a Linux desktop, with a Nvidia Titan Z GPU (for deep learning). The deep learning framework utilized was Keras/Theano.
Promising results were obtained. While the CNN performs most of the heavy lifting, the addition of spatial features extracted from the landmark detection algorithm further improved results, and also resulted in reducing potential over-fitting (reduced over-fitting on pixel intensity). A SVM approach to classification was utilized over the standard Softmax classification in CNN, which also slightly improved results.
Publications:
Roth H, Lu L, Farag A, Sohn A, Summers RM. Spatial aggregation of holistically-nested networks for automated pancreas segmentation. Medical Image Computing and Computer Assisted Intervention, 2016 (in press).
Sohn A, Roth H, Summers RM. Deep Organs: Towards automated abdominal organ segmentation using convolutional neural networks and landmark detection. Medical Image Computing and Computer Assisted Intervention, 2016 (in progress).