Computational Decision Support Systems for Cancer Pathology Are Becoming More Widespread

On July 15, 2019, researchers from Memorial Sloan Kettering Cancer Center and the Department of Pathology at the Weill Cornell Graduate School of Medical Sciences, led by Thomas J. Fuchs, presented in Nature Medicine a learning-based system using deep learning (deep learning). Capable of training accurate classification models on an unprecedented scale, the system laid the foundation for deploying computational decision support systems in clinical practice.

The authors reported that developing decision support systems for pathology and using them in clinical practice had previously been impractical because of the need for large, manually annotated datasets.

The deep learning system they describe overcomes this problem. In particular, the new system uses only reported diagnoses as labels for training, avoiding expensive and time-consuming manual annotations at the pixel level.

The team scaled this framework to a dataset of 44.732 whole-slide images from 15.177 patients, without any data curation.

The System Was Able to Diagnose Prostate Cancer, Basal Cell Carcinoma of the Skin, and Axillary (Armpit) Metastasesof Breast Cancer With 98% Accuracy.

When applied clinically, the system allowed pathologists to avoid examining 65-75% of pathology blocks while maintaining 100% sensitivity.

The authors concluded that their results showed the system’s ability to train accurate classification models, laying the foundation for deploying computational decision support systems in clinical practice.

For those working in medical informatics, the publicly shared MSK breast cancer metastasis dataset can be accessed here. The source code for this study is available for download here .

• All Our News on Artificial Intelligence in Cancer