Deep Domain Adaptation Learning Framework for Associating Image Features to Tumour Gene Profile

Abstract

While medical imaging and general pathology are routine in cancer diagnosis, genetic sequencing is not always assessable due to the strong phenotypic and genetic heterogeneity of human cancers. Image-genomics integrates medical imaging and genetics to provide a complementary approach to optimise cancer diagnosis by associating tumour imaging traits with clinical data and has demonstrated its potential in identifying imaging surrogates for tumour biomarkers. However, existing image-genomics research has focused on quantifying tumour visual traits according to human understanding, which may not be optimal across different cancer types. The challenge hence lies in the extraction of optimised imaging representations in an objective data-driven manner. Such an approach requires large volumes of annotated image data that are difficult to acquire. We propose a deep domain adaptation learning framework for associating image features to tumour genetic information, exploiting the ability of domain adaptation technique to learn relevant image features from close knowledge domains. Our proposed framework leverages the current state-of-the-art in image object recognition to provide image features to encode subtle variations of tumour phenotypic characteristics with domain adaptation techniques. The proposed framework was evaluated with current state-of-the-art in: (i) tumour histopathology image classification and; (ii) image-genomics associations. The proposed framework demonstrated improved accuracy of tumour classification, as well as providing additional data-derived representations of tumour phenotypic characteristics that exhibit strong image-genomics association. This thesis advances and indicates the potential of image-genomics research to reveal additional imaging surrogates to genetic biomarkers, which has the potential to facilitate cancer diagnosis.