Semisupervised Ensemble CNNs for Computer Aided Diagnosis

David Chapman, Assistant Professor of Computer Science
Yelena Yesha, Professor of Computer Science
Dorsa Zaiei, Graduate Research Assistant
Sumeet Menon, Graduate Research Assistant

Our goal is to evaluate semisupervised deep learning models for Computer Aided Diagnosis of radiology imagery from a representative sample of radiology reports. UMBC will be receiving a sample of 3 million images and associated radiology reports from partners with Mercy hospital. Deep Learning models have achieved state of the art performance on a variety of medical imaging diagnostic tasks in recent years with a few top performing models achieving accuracy comparable to human pathologists for specific tasks. However, the need for pixel level classification labels by expert medical imagers is a tedious, challenging, and even potentially contentions process that limits that size and diversirty of training data for Fully Supervised techniques. Multiple Instance Learning (MIL) and related methods using Expectation Maximization (EM) can be used to automatically learn distinguishing convolutional features using image-level radiology reports presentl!
y available. The HPCF resources will allow us to train these semisupervised model over this large dataset efficiently by making use of parallel GPU resources thereby greatly reducing the computational time necessary to develop the intended pixel classifier for Computer Aided Diagnosis as well as enable research in the training of application of ensemble CNNs to further improve performance without compromising training accuracy.