All Times ET
Keywords: Convolutional Neural Networks, Generative Adversarial Networks, Explainable AI, Radiology
Convolutional neural networks (CNNs) have become valuable instruments for computer vision in medical imaging, able to learn features of disease without explicit programming. However, algorithm transparency is necessary for these to be applied in clinical practice. To address this, we propose a feature interpretation generative adversarial network (FIGAN) to generate synthetic images that facilitate CNN feature interpretation. Feasibility of the proposed approach was assessed on a previously-developed CNN designed to assess contrast enhancement adequacy of liver MR images for lesion detection. Review of FIGAN images revealed that this CNN utilizes features related to tissue-vessel contrast, nodular liver texture, and tissue brightness to determine adequacy of contrast enhancement.