Machine learning and neural networks continue to prove effective tools to improve healthcare. In 2017 we wrote about a Stanford University study in which a deep convolutional neural network performed comparably to 21 board-certified dermatologists. A large group of German researchers associated with the German Cancer Research Center, Heidelberg University, and other institutions went further to test deep learning against pathologists in detecting histopathological images of melanoma. The study was published in the September 2018 issue of the European Journal of Cancer.
In ordinary practice, when a dermatologist suspects a mark or growth on a patient’s skin may be cancerous, the physician removes a small sample of tissue and sends it to a pathology laboratory for a biopsy. Lab personnel capture and examine histopathological images of suspect tissue with a microscope. According to the research team, there is a significant 25% to 26% difference of opinion among histopathologists in classifying benign tissue from malignant melanoma.
To compare deep learning to the performance of trained histopathologists, the German researchers started by training a convolutional neural network (CNN) with 595 tissue images, half of which were benign and half were malignant. The training images were part of a large group of 695 images classified by an expert histopathologist.
The remaining 100 images were used to test the neural network and 11 certified histopathologists. Even with the limited information available, the CNN consistently outperformed the histopathologist in classifying melanoma images.
The German research does not suggest replacing histopathologists with neural networks. The healthcare professionals can reference a much wider range of information about individual patients before making a final determination on a given tissue sample. The study does suggest, however, the value of using deep learning to supplement human assessment for a final melanoma diagnosis.
