Association of Pathology Chairs: Abstracts of the 2019 Annual Meeting: Innovation Through Collective Excellence: Shaping the Future of Pathology July 21 to 24, 2019, Boston Seaport Hotel and World Trade Center, Boston, MA. (2019). Academic Pathology. https://doi.org/10.1177/2374289519852559
The previously published abstract for APC-19-0014PO had the correct title, but incorrectly duplicated the abstract content for APC-19-0006PO. The corrected abstract and author list for APC-19-0014PO are published below.
APC-19-0014PO. Utility of Artificial Intelligence (AI) / Machine Learning (ML) in the Histology Education Arena
E. Vali Betts1, A. Dubrovsky1, K. Olson1, K. Beck1, A. Tsirigos2, V. Harnik2, A. Galvao Neto2, E. Adler2, and H. Rashidi
1
1University of California, Davis, Sacramento, CA, USA
2New York University School of Medicine, New York, NY, USA
Objectives: Recognizing and differentiating the histologic features of various tissue types can be a challenging task for new learners and understanding the subtle variations in the histology look-a-likes is an essential component of building a strong foundation in this discipline. Building an AI/ML tool that can identify and generate a differential diagnosis for various tissue histology can assist students in their learning process. In this study we are generating multiple models, to identify the most accurate ML platform that can perform the aforementioned tasks. Methods: Multiple ML models are being generated for identifying histologic entities and their respective look-a-likes. The training set includes images from multiple databases (acquired from University of California Davis and New York University’s histology digital teaching and research slide databases respectively). Approximately 1000 images (20 images at 4x and 10x magnifications on approximately 50 categories) are utilized in these training sets to build the appropriate models. A transfer learning approach on two well established deep convolutional neural networks (ResNet50 and SqueezeNet) is being employed to generate these models. The most accurate models that are also able to generate the appropriate differential diagnosis are then deployed in an App with a friendly user interface. Results: Our preliminary results on models built with a subset of the histology categories (35 categories) shows an accuracy of 91.3% in correctly identifying the tissue subtype. The full study that includes both institution’s digital databases is underway. Conclusions: These ML tools will not replace the current approach in histology training but rather assist learners with certain aspects of this educational journey. Such tools may also be invaluable in underserved areas where educational resources are sparse, especially since the application platform (Core-ML, etc.) that it is being deployed on is independent of internet and wireless services.
