Study design

The overall goal of our study was to demonstrate where artificial intelligence technologies could provide efficiency gains for intraoperative margin assessment by: (1) generating a 3-dimensional model of tissue for grossing recommendations, (2) rapid localization of remaining tumor, and (3) mapping histological findings to a display output familiar to the surgeon. We first compared tissue gross measurements (length, width, height) to computer generated measurements. Then, we calculated AUCs to communicate the accuracy of a CNN-GNN for tumor localization. Finally, the surgeon reported the number of cases which exhibited concordance between hand-drawn and computer-generated tumor maps at the correct anatomical position and orientation. Throughout the study, we consulted with end-users (pathologists, surgeons) for the design of the digital displays, and we utilized an independent test set of randomly selected cases. Patients were randomized into training, validation and held-out test set cohorts (i.e., serial sections and WSI subimages were included in the same cohort to avoid target leakage and inflating test set statistics). Sample size was based on data availability to calculate concordance and accuracy. Confidence intervals for the study findings were calculated using non-parametric bootstrapping and posterior distributions. The authors complied with all relevant ethical regulations including the Declaration of Helsinki. Human Research Protection Program (institutional review board, IRB) of Dartmouth Hitchcock Medical Center gave ethical approval for this work. All necessary patient/participant consent was obtained, including written consent, and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals.