PROJECT SUMMARY/ABSTRACT There has been significant work in creating tools that leverage computer vision algorithms to automate medical image analysis. Most of these algorithms have been developed for natural images, which are usually single static images that can be treated individually. However, medical images are usually part of a study that may include various views and orientations that are considered together with other clinical data when making a diagnosis. Three dimensional convolution neural networks (CNN) can address this issue in part when images are evenly spaced, but many medical imaging modalities such as ultrasound (US), fluoroscopy, and biopsy imaging have variable orientations and irregular spacing. Graph convolutional networks (GCN) have the potential to address this issue as they generalize the assumptions of CNNs to work on arbitrarily structured graphs. Automatic thyroid nodule detection in ultrasound (US) is one application that such a graph-based approach could have a large impact. The thyroid cancer incidence rate has tripled in the past thirty years, with an estimated cost of $18-21 billon in 2019. US is the imaging modality of choice, which consists of multiple 2D images of different locations and orientations. US readings are often vague and subjective in nature, which has resulted in a steady increase in the number of biopsies performed over the past 20 years. It is estimated that about one-third of all thyroid biopsy procedures performed in the United States are medically unnecessary, leading to the unmet need for noninvasive diagnostic tests that can reliably identify which nodules require a biopsy. The research objective of this R21 is to develop a new graph-based approach to leverage spatial information contained within imaging studies that will be combined with biomarkers and other known risk factors. Our graph model will enable more complete detection of thyroid cancer, as well as the prediction of future cancer aggression, both with spatially localized explanations. GCN features will be used to predict voxel-level cancer suspicion, thereby enabling a novel method for performing “imaging biopsy.” Finally, voxel-level suspicion maps will be aggregated into patient-level quantitative imaging biomarkers and combined with clinical data to create a multimodal nomogram for performing risk stratification.