Fine-tuning.

Top Line: Prognosis for colorectal cancer falls along a far-reaching spectrum.

The Study: In this study, specimens from nearly 2500 colorectal cancer patients from four large Norwegian and UK cohorts were used to successively train and tune a deep learning system for predicting outcomes for stage I-III colorectal cancer patients from simple H&E slide input data. The system was initially trained on specimens with clear “good” or “bad” outcomes, which together comprised one-third of the training cohort. A “good” outcome was defined as no cancer recurrence or cancer death 6 years after surgery. “Bad” was defined as cancer death between 100 days to 2.5 years after surgery. Anything in-between was considered a “non-distinct” outcome and was used to tune the system. Data (i.e. slide) preparation was done at multiple centers (potentially improving robustness) and included both 10x and 40x resolution images. Next, a segmentation algorithm was trained to identify tumor tissue. In the critical step, 10 separate neural networks (5 at low resolution, 5 at high resolution) were trained to predict outcomes. The outcomes from each ensemble of neural networks were averaged, and the degree of agreement determined the overall DoMore-v1-CRC classifier score. That score indeed proved a strong predictor of cancer-specific survival in the validation cohort, even in multivariable analysis. It also performed well among different stages. Accuracy was 76%. Sensitivity and specificity were 52% and 78%, respectively, and PPV and NPV were 19% and 94%.

TBL: A deep learning classifier that uses simple H&E slides may one day help us decipher bad colorectal cancers. | Skrede, Lancet 2020