Title: Deep learning applications in the field of pathology

Biography:

Jerome Cheng is a Clinical Assistant Professor of pathology in the Division of Pathology Informatics, in the Department of Pathology at University of Michigan, USA. He had his residency training in anatomic and clinical Pathology at SUNY Health Science Center in Brooklyn, New York and fellowship in pathology informatics at University of Michigan. He is board certified in AP/CP and Clinical Informatics. His research interests include image analysis, natural language processing, data mining, and application of machine learning techniques such as convolutional neural networks and Random Forests to image and non-image based medical datasets.

Abstract

In recent years, convolutional neural networks (CNN)  have revolutionized the field of computer aided digital image analysis due to their accuracy in classifying and recognizing objects /patterns from different subject matters, including histopathological images. Aside from object classification, CNN is also useful for object detection/tracking and semantic segmentation. There is a lot of potential for deep learning to change and improve the practice of pathology due to the highly visual nature of the field, especially in tasks that pathologists find boring and tedious such as finding mitosis and counting cells. Other example applications of CNN in pathology include classification of lesions (e.g. benign vs. malignant), cancer localization, nuclei identification, and segmentation of regions of interests in an image. In the image below, a CNN model based on U-Net, together with a ground truth map, was able to accurately highlight regions of glomeruli in a kidney section. Running CNN experiments used to require a significant amount of programming expertise. Fortunately, open-source tools that need minimal or no programming knowledge such as Tensorflow and Orange (Biolab) are now available, making deep learning accessible to a wider audience, including Pathologists.