Fig. 2

Algorithm schema. Training and inference pipeline for slice-assisted nodule detection. Training consists of using 2-dimensional axial and coronal MIP slices from each CT being input into a Retinanet model. Inference adds three additional steps to the Retinanet raw inferences: unsupervised clustering, false positive reduction using clustering metadata (max and mean confidence scores, whether the cluster contains both axial and coronal raw inferences, the number of raw inferences clustered together, and distance from top and bottom of the CT in mm)