Abstract
Generalizability of deep-learning (DL) model performance is not well understood and uses anecdotal assumptions for increasing training data to improve segmentation of medical images. We report statistical methods for visual interpretation of DL models trained using ImageNet initialization with natural-world (T (II)) and supervised learning with medical images (L (MI)) for binary segmentation of skin cancer, prostate tumors, and kidneys. An algorithm for computation of Dice scores from union and intersections of individual output masks was developed for synergistic segmentation by T (II) and L (MI) models. Stress testing with non-Gaussian distributions of infrequent clinical labels and images showed that sparsity of natural-world and domain medical images can counterintuitively reduce type I and type II errors of DL models. A toolkit of 30 T (II) and L (MI) models, code, and visual outputs of 59,967 images is shared to identify the target and non-target medical image pixels and clinical labels to explain the performance of DL models.