Lancet Oncol: Deep learning models that predict microsatellith instability with HE staining slices!

Detecting colorectal cancer microsatellithasal instability (MSI) is critical to clinical decision-making because it distinguishes between patients who predict different treatment responses and prognosis.
recommended universal MSI testing for colorectal cancer patients, but many patients have not yet been tested.
urgent need for widely available, cost-effective tools to help patients choose to test.
rikiya and others have studied a deep-learning-based system for automatically predicting the potential of MSI directly from the entire slice map (WSIS) dyed by sumu Jingyi (H.E.).
deep learning model (MSINet) was based on 100 WSIS (50 microsatellith stabilization (MSS) and 50 MSIs that were dyed 40 times larger.
internal validation was performed in the retention test set (15 cases of H&E staining WSIS; 7 cases of MSS and 8 cases of MSI) and external verification was performed in 484 cases of H.E. staining WSIS (402 cases of MSS and 77 cases of MSI;479 patients) in the TCGA database.
the predictive effect based on sensitivity, specificity, negative prediction value (NPV) and area under the subject's working characteristic curve (AUROC).
randomly selected a subset of WSIS (20 each for MSS and MSI) from an external data set to compare the model's predictions with 5 gastrointestinal pathologists.
MSINet's performance in the TCGA-CRC dataset is 0.931 (95% CI 0.771-1.000) on the reserved test set from the internal dataset and 0.779 (0.720-0.838) on the external dataset.
in the external data set, using sensitivity-weighted operating points, the model's NPV can reach 93.7% (95% CI 90.3 to 96.2), sensitivity is 76.0% (64.8 to 85.1), specificity is 66.6% (61.8 to 71.2).
msINet model compared to the performance of five pathologists in the reader experiment (40 cases), the AROC of the model was 0.865 (95% CI 0.735-0.995).
average AUROC score of 0.605 (0.453-0.757) for the five pathologists in the united States.
, the deep learning model predicted that MSI outpernumbered experienced gastrointestinal pathologists based on WSIS dyed by H. E.
in the current msI testing model, the model may be used as an automated screening tool to validate patients, reducing unnecessary testing and saving significant testing-related labor and costs.
。