Creating the Black Box: A Primer on Convolutional Neural Network Use in Image Interpretation

Abstract

Convolutional neural networks have been shown to demonstrate high diagnostic performance in radiologic image interpretation tasks ranging from recognition of acute stroke on computed tomography to identification of tuberculosis on plain radiographs. To a radiologist not immersed in computer science jargon, it may seem that this inscrutable black box is best treated warily, at arm's length. In this work, we illustrate how a radiologist without a deep background in computer science may be able to set up a state-of-the-art convolutional neural network for image interpretation tasks through transfer learning. This technique is relatively simple to implement, has been shown to demonstrate equivalent performance to neural networks specifically trained on medical image data, and offers a chance for the interested-but-intimidated radiologist to deep her toe in the water without becoming overwhelmed.