Murdoch University Research Repository

Welcome to the Murdoch University Research Repository

The Murdoch University Research Repository is an open access digital collection of research
created by Murdoch University staff, researchers and postgraduate students.

Learn more

Translating AI to digital pathology workflow: Dealing with scarce data and high variation by minimising complexities in data and models

Somaratne, Govini Tantrige Upeka Vianthi (2021) Translating AI to digital pathology workflow: Dealing with scarce data and high variation by minimising complexities in data and models. PhD thesis, Murdoch University.

PDF - Whole Thesis
Download (13MB) | Preview


The recent conversion to digital pathology using Whole Slide Images (WSIs) from conventional pathology opened the doors for Artificial Intelligence (AI) in pathology workflow. The recent interests in machine learning and deep learning have gained a high interest in medical image processing. However, WSIs differ from generic medical images. WSIs are complex images which can reveal various information to support different diagnosis varying from cancer to unknown underlying conditions which were not discovered in other medical investigations. These investigations require expert knowledge spending a long time for investigations, applying different stains to the WSIs, and comparing the WSIs. Differences in WSI differentiate general machine learning methods that are applied for medical image processing. Co-analysing multistained WSIs, high variation of the WSIs from different sites, and lack of labelled data are the main key interest areas that directly influence in developing machine learning models that support pathologists in their investigations. However, most of the state-ofthe- art machine learning approaches cannot be applied in the general clinical workflow without using high compute power, expert knowledge, and time. Therefore, this thesis explores avenues to translate the highly computational and time intensive model to a clinical workflow. Co-analysing multi-stained WSIs require registering differently stained WSI together. In order to get a high precision in the registration exploring nonrigid and rigid transformation is required. The non-rigid transformation requires complex deep learning approaches. Using super-convergence on a small Convolutional Neural Network model it is possible to achieve high precision compared to larger auto-encoders and other state-of-the-art models. High variation of the WSIs from different sites heavily effect machine learning models in their predictions. The thesis presents an approach of using a pre-trained model by using only a small number of samples from the new site. Therefore, re-training larger deep learning models are not required which saves expert time for re-labelling and computational power. Finally, lack of labelled data is one of the main issues in training any supervised machine learning or deep learning model. Using a Generative Adversarial Networks (GAN) is an approach which can be easily implemented to avoid this issue. However, GANs are time and computationally expensive. These are not applicable in a general clinical workflow. Therefore, this thesis presents an approach using a simpler GANthat can generate accurate sample labelled data. The synthetic data are used to train classifier and the thesis demonstrates that the predictive model can generate higher accuracy in the test environment. This thesis demonstrates that machine learning and deep learning models can be applied to a clinical workflow, without exploiting expert time and high computing power.

Item Type: Thesis (PhD)
Murdoch Affiliation(s): IT, Media and Communications
Supervisor(s): Wong, Kevin, Laga, Hamid and Wang, Alex
Item Control Page Item Control Page


Downloads per month over past year