Dealing with Multi-Dimensional Data and the Burden of Annotation

• Published in: The American journal of pathology Vol. 191; no. 10; pp. 1709 - 1716
• Main Authors: Mitchell, Benjamin R., Cohen, Marion C., Cohen, Stanley
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.10.2021
• ISSN: 0002-9440; 1525-2191
• DOI: 10.1016/j.ajpath.2021.05.023

The need for huge data sets represents a bottleneck for the application of artificial intelligence. Substantially fewer annotated target lesions than normal tissues for comparison present an additional problem in the field of pathology. Organic brains overcome these limitations by utilizing large numbers of specialized neural nets arranged in both linear and parallel fashion, with each solving a restricted classification problem. They rely on local Hebbian error corrections as compared to the nonlocal back-propagation used in most artificial neural nets, and leverage reinforcement. For these reasons, even toddlers are able to classify objects after only a few examples. Rather than provide an overview of current AI research in pathology, this review focuses on general strategies for overcoming the data bottleneck. These include transfer learning, zero-shot learning, Siamese networks, one-class models, generative networks, and reinforcement learning. Neither an extensive mathematic background nor advanced programing skills are needed to make these subjects accessible to pathologists. However, some familiarity with the basic principles of deep learning, briefly reviewed here, is expected to be useful in understanding both the current limitations of machine learning and determining ways to address them.