Skip to main content
. Author manuscript; available in PMC: 2021 May 5.
Published in final edited form as: Mod Pathol. 2020 Dec 10;34(5):862–874. doi: 10.1038/s41379-020-00724-3

Figure 1. Deep learning experimental strategy.

Figure 1.

A. Distribution of thyroid neoplasm subtypes within the institutional (University of Chicago Medicine, UCM) and TCGA datasets. The institutional dataset contained a total of 115 slides across four subtypes, as detailed here. The TCGA thyroid cohort (THCA) contained 497 slides across three subtypes: papillary thyroid carcinoma, not otherwise specified (PTC-NOS), papillary thyroid carcinoma, follicular variant (PTC-FV), and papillary thyroid carcinoma, columnar cell variant (PTC-tall). As the annotated subtypes in TCGA use an older naming convention before the recognition of NIFTP, these annotations cannot be directly compared between datasets. B. Cross-validation plan for training a deep learning model. A deep learning model was first trained on the institutional dataset to predict tumor subtype, using three-fold cross-validation as illustrated here. Patients were separated into three categories of 39, 38, and 38 patients per category, with annotated subtypes balanced between categories. For each of three iterations, two-thirds of the dataset was used for training and one-third was used for validation to generate predictions. Predictions in each of the left-out validation sets were aggregated across the three k-fold iterations in order to generate the reported cross-validation performance statistics. C. Overview of the deep learning training strategy used to generate cross-validated subtype predictions on the institutional dataset, as also described in B. D. After cross-validation was completed, a final model (DL-UCM-ST) was trained across the entire institutional dataset to predict subtype. This model was applied to the TCGA dataset, for which subtype predictions were generated. E. Deep learning models were trained on the TCGA dataset to predict BRS, using cross-validation as described in B. F. After cross-validation was completed, a final model (DL-TCGA-BRS) was trained across all TCGA slides to predict BRS. This model was then used to generate BRS predictions on the institutional dataset.