Fig. 7: Uncertainty thresholding in a synthetic test using GAN-generated images.

a A class-conditional generative adversarial network (GAN) was trained on TCGA to generate adenocarcinoma (LUAD) or squamous cell carcinoma (LUSC) synthetic images. Using embedding interpolation, intermediate neutral images are also generated to approximate images near the decision boundary. Example synthetic images are shown here using the LUAD, LUSC, and Intermediate class labels. b Using a model trained on the full TCGA dataset, predictions were calculated for 1000 LUAD, LUSC, and Intermediate GAN images. Synthetic LUAD and LUSC images were predicted accurately, and Intermediate synthetic images showed an even spread of predictions. c Models were trained with the addition of varying amounts of GAN-Intermediate slides with randomly assigned labels. Cross-validation slide-level area under receiver operator curve (AUROC) is shown. Performance degrades with increasing proportion of GAN-Intermediate slides. The shaded intervals represent the AUROC 95% confidence interval at each dataset size. d Cross-validation slide-level AUROC is shown from models trained with a dataset size of 500 slides plus varying amounts of GAN-Intermediate slides. Performance degrades as increasing number of uninformative GAN-Intermediate slides are added, but performance in the high-confidence uncertainty quantification (UQ) cohorts remains high despite large numbers of uninformative slides. For the +0%, +10%, +20%, +30%, +40%, and +50% GAN experiments, p values comparing high-confidence AUROC to AUROC without UQ are 0.00032, 0.00014, 0.0034, <0.0001, 0.00068, and 0.00090, respectively. Statistical comparisons were made with one-sided, paired t-tests. e Distribution of low- and high-confidence predictions in the experiments shown in d. Virtually none of the GAN-Intermediate slides are classified as high-confidence in these experiments. For all boxplots, center line represents the median (50th percentile), lower and upper box bounds represent interquartile range (25th–75th percentile), and minimum (lower whisker) and maximum (upper whisker) bounds extend to furthest datapoint up to 1.5 times the interquartile range, with outliers shown as diamonds. Source data are provided as a Source Data file.