PhD Defense by Amal Lahiani

Speaker: Amal Lahiani Date: Thursday, October 22, 2020 Update: We want to congratulate Amal to the very successful defense of her PhD thesis.

Deep Learning Solutions for Cancer Drug Development in Digital Pathology

Abstract:
In the domain of oncology, pathology tasks such as patient diagnosis and cancer drug development, have been revolutionized since the growing availability and quality of high-resolution whole slide scanners. The transition from conventional glass slide microscopic assessment to digital pathology has been beneficial and promising for both pathologists and researchers due to the ability to collect, process and analyze much larger datasets with more stains (clinical markers) using more objective, accurate and consistent methods. Furthermore, with the recent advances in computational power, deep learning solutions have been developed allowing to explore previously unachievable predictive algorithms and to enhance the value and insights that can be generated from pathology tissue images. However, different challenges need to be taken into consideration in order to make the applications more efficient and to increase confidence and adoption in the very conservative field of human pathology. These challenges include, among others, application generalization over multiple stains, limitations in terms of tissue staining materials and procedures, size of histologic whole slide images (WSis) and intra and inter-pathologist variability. In this thesis, we explore new methodologies, techniques, and deep learning solutions to the aforementioned challenges in the context of two different applications: stain generalization and stain virtualization applied to digital images of Colorectal Carcinoma metastases in liver tissue from biopsy and surgical specimen.

Slide annotation is a key step in pathology routines and in cancer biomarker research aiming to quantify pattern changes in microscopic WSis of tumor biopsies. The process of manual annotation can be tedious and subjective, especially in the context of drug development research where a correlative analysis of multiple biomarkers is required. In the first part of this dissertation, we elaborate on a supervised learning-based method that allows to segment different functional compartments in histology images of various stainings including Hematoxylin and Eosin (H&E) staining and multiple immunohistochemistry (IHC) stainings. We show the effect of the proposed solution on increasing the stain generalization performance and dealing with dataset variability by comparing to state-of-the-art methods and using different visualization techniques.

In the second part, we propose stain virtualization solutions allowing to virtually generate an IHC staining from different input stainings using unsupervised learning methods. Stain virtualization presents many advantages to the pathology drug development workflow, such as reduction of lab workload, reduction in tissue and costs, and multiplexing of different biomarkers on the same tissue and with the same coordinate system. In this context, we additionally propose solutions to one of the main challenges of high-resolution style transfer in general and to WSI processing in particular, i.e. the tiling artifact caused by the necessity of tilewise processing. We describe different experiments and visualizations aiming to prove the hypothesis and we validate the application mathematically and histologically.