Artificial Intelligence (AI) and Machine Learning (ML) are dependent upon a large quantity of image sets and a high quality of tagging those image sets. The majority of data sets used to train algorithms in the industry use Natural Language Processing (NLP) to search for key phrases within radiology reports and use those phrases to tag images with findings. This silver-standard of image tagging results in a lower quality of recognition algorithms as some images are tagged incorrectly. IDgital creates gold-standard image tagging by having key image objects created by the physician at the time of interpretation. This gold-standard tagging ensures a high quality and accurate data set. In addition, when an algorithm identifies a particular finding, the physician’s interpretation results in a human feedback loop resulting in enhancements to the algorithms. In addition, by leveraging the Google Cloud Platform (GCP), our ML is supported by Google’s proprietary Tensor Processing Units (TPU) which were specifically designed for neural network machine learning.
You may have noticed many capchas on websites asking you to identify store fronts or street signs. AI can easily identify a sign, but knowing what is a street sign for traffic as opposed to a sign advertising a store is a much more difficult process. By having humans identify street signs vs store signs, you are participating in this same type of human feedback loop to help enhance these algorithms that will be used in self-driving cars.