Multi-task Feature Learning to train Interpretable Encoders on Histopathology Images. And Epigenomic Oncofetal Reprogramming Across Cancer Types

About this Class

For our next CDSL webinar we will have two fellows’ talks from Dr. Zisha Zhong and Arati Rajeevan. Zisha’s talk details: Title: Multi-task Feature Learning to train Interpretable Encoders on Histopathology Images Digital pathology images contain very detailed information of tumor micro-environment (TME) at a micrometer resolution. Existing research on digital pathology images focuses on either cancer diagnosis or prognosis, or morphological segmentation. However, so far as we know, little frameworks exist to perform differential analysis between images from diverse groups. We developed a multitask multi-instance framework for differential analysis of digital pathology images of breast cancer groups. We have developed a Multitask framework to learn low-dimensional representations of Hematoxylin and Eosin (H&E) images at patch-level and image-level. We utilize a convolution neural network (CNN) to obtain a patch-level representation and a gated-attention mechanism to aggregate patch-level representations to get an image-level representation. Multitask-HENet is trained by predicting numerous image-level annotations, including survival outcome, stage, molecular and immune subtype, gene expression, cytotoxic lymphocytes level, and tumor immune dysfunction and exclusion scores under a multitasking framework. We train the model on TCGA-BRCA H&E images and annotations along with the patient-matched annotations from other studies. By performing a diverse set of image-level tasks through a shared patch-encoder, the feature network exhibits the distinction among tumor, stroma, lymphocyte, necrosis, fat, and other tissue types. Bio: Zisha Zhong is a postdoctoral fellow in the Cancer Data Science Laboratory (CDSL), National Cancer Institute (NCI). Under the supervision of Dr. Peng Jiang, Zisha mainly focuses on understanding biomedical images using machine learning approaches. He received a B.E. degree from Central South University (CSU) in 2010 and a Ph.D. degree from the National Laboratory of Pattern Recognition (NLPR), Institute of Automation, Chinese Academy of Sciences (CASIA) in 2017 under the supervision of Dr. Chunhong Pan and Dr. Bin Fan. From March 2017 to July 2018, he worked as a Postdoc on medical image analysis at the University of Iowa Health Care (UIHC) under the supervision of Dr. Xiaodong Wu and Dr. Yusung Kim. He has broad research interests in pattern recognition, machine learning, and image analysis.   Arati’s talk details: Title: Epigenomic Oncofetal Reprogramming Across Cancer Types Developmental biologists have long hypothesized that tumorigenesis involves a reactivation of embryonic developmental programs that would normally be dormant in healthy tissues. For many years it has been difficult to test this hypothesis due to the lack of genomic sequencing data across tumor and tissue types, but through recent technological advances it is now possible. To answer the question of whether genomic regions necessary in development are re-activated in cancer, I obtained epigenomic and transcriptomic data from several tissues and selected genomic regions based on trends in their expression profiles across fetal, adult, and tumor tissues. I found that the genes that are active in both development and cancer are prognostic indicators of survival, have been experimentally proven to increase cell proliferation, and are functionally linked to enhancers that are themselves re-activated. These genes make attractive targets for cancer therapies, as altering their function may reduce the severity of different types of cancers. Bio: Arati is a post-baccalaureate fellow in Dr. Sridhar Hannenhalli’s lab in the CDSL. She graduated from Carnegie Mellon University in 2019 with a B.S. in Biological Sciences. Her research interests include exploring the effect of CREs and other noncoding regions on cancer progression and better understanding the origins of cancer to improve diagnostics and treatment.