Supplementary Materials

The PDF file includes:

  • Materials and Methods
  • Fig. S1. t-SNE plot for Mirai’s hidden representation (left) without and (right) with adversarial training on 5000 random samples from the MGH test set.
  • Fig. S2. Saliency scores of images and all clinical risk factors across the MGH test set.
  • Fig. S3. t-SNE plots for Mirai’s hidden representation colored by cancer subtype factors on 1000 random positive examinations from the Karolinska test set.
  • Table S1. The distribution of clinical risk factors in the MGH dataset.
  • Table S2. ROC AUCs and C-indices for Mirai and prior risk models on all test sets excluding cancers confirmed within 6 months of the screening mammogram.
  • Table S3. Ablation study of Mirai on the MGH datasets.
  • Table S4. C-index for different models on different subpopulations in the MGH test set.
  • Table S5. C-indices and ROC AUCs for Mirai in predicting cancers of different subtypes in the Karolinska test set.
  • Table S6. Number of examinations per cancer type in the Karolinska dataset.
  • Table S7. Sensitivity and specificity of different risk models in identifying high-risk cohorts at MGH, excluding mammograms with a BI-RADS 0 assessment that were followed by a cancer diagnosis within 1 year.
  • Table S8. Distribution of follow-up times and times until cancer diagnosis for examinations in the MGH, Karolinska, and CGMH test sets.
  • References (6372)

[Download PDF]

Other Supplementary Material for this manuscript includes the following:

  • Data file S1 (Microsoft Excel format). Primary data from figures.