Radiology and digital pathology images are used to examine different regions of the human body to identify and characterise possible pathology and to assist in diagnosis. These imaging procedures generate large volumes of sophisticated digital imaging data which can overwhelm medical professionals in systematically analysing and interpreting. In this direction, machine learning and deep learning algorithms coupled with radiomics analysis, are rapidly proving to be the state-of-the-art foundation, achieving enhanced performances in various medical image analysis problems such as detection, segmentation, computer-aided diagnosis, non-invasive molecular characterisation of different tissue types and associating morphologies observed in biomedical images with clinical outcome and genomics for several diseases. However, development of such deep learning-based predictive and correlation models would involve addressing several challenges in both radiology and digital pathology image analysis. The thesis identifies some of these challenges and discusses the proposed methodologies for addressing the same. The thesis can be seen in the light of two central applications of deep learning techniques, namely- automated analysis and knowledge discovery in radiology and digital pathology.

Deep learning is a subset of machine learning algorithms that can learn and improve from experience processing the input data. Recent technical advances in deep learning neural network architectures have led to significant progress in automated analysis of various forms of data, such as text, image and speech. The core aspect for these advances in deep learning is its ability to discover hierarchical feature representations solely from the input data automatically. Deep learning approaches significantly contrasts with other classical machine learning-based techniques, which primarily focus on hand-crafted feature engineering based on domain-specific knowledge.

Deep fully convolutional neural network-based architectures have shown great potential in medical image segmentation. However, such architectures usually have millions of parameters, and an inadequate number of training samples leads to overfitting and poor generalisation. The thesis has attempted to address these limitations by proposing a novel fully convolutional multi-scale residual DenseNets architecture and a customised cost function for addressing class imbalanced datasets. The proposed architecture applied to MRI cardiac multi-structures segmentation showed performance on par with the state-of-the-art approaches and achieved a parameter reduction factor of 100 times compared to other methods. The average attainable Dice score for segmentation of MRI cardiac multi-structures in short-axis view has been 0.90 for the left ventricle, right ventricle and myocardium. From the segmentation maps, clinically relevant cardiac parameters and hand-crafted features which reflect the clinical diagnostic analysis are extracted, and an ensemble classification system is trained for cardiac disease classification. The proposed methodology for automated cardiac disease diagnosis task achieved near-perfect classification.

The spatial resolution of whole slide images (digital pathology) is extremely high and hence direct application of CNN based approaches for semantic segmentation or classification is infeasible because of the hardware limitation. Hence, the proposed methodology is to divide the whole slide image into patches of computationally feasible resolution and analyse them independently. However, analysing whole slide images with patch-based approaches necessitates sampling best representative patches for effectively training deep learning models. One of the critical areas of research is developing effective patch sampling strategies for whole slide images when the provided annotations in the training data are at pixel-level and slide-level. The thesis addresses these issues by developing an ensemble deep learning-based framework for automated tumour segmentation, and downstream analyses of H\&E stained whole slide images when pixel-level annotations are provided. The performance of the proposed framework for three cancer sites (detection of lymph node metastases in breast cancer and pN-staging, segmentation of viable tumour region in liver cancer tissue and lesion segmentation in colon cancer tissue) was on par with state-of-the-art methods. The thesis attempted to address the problem of classification of whole slide images when only slide-level annotations are provided by proposing a deep learning-based framework for unsupervised discovery of representative patch regions. The proposed methodology for automated histologic subtype classification of low-grade gliomas achieved an accuracy score of 80%.

Histopathology and molecular characterisation of a tumour require acquiring biopsy, and the reliability of biopsy is constrained by spatial location. Additionally, the complications associated with the biopsy procedure itself limits its usage. Quantitative imaging features derived from the radiomics analysis have the potential to predict the molecular characteristics of tumours non-invasively. The thesis examines the utility of hand-crafted and end-to-end deep learning-based radiomics analysis for the task of predicting histological subtype of low-grade gliomas and molecular subtype of glioblastoma multiforme respectively using pre-operative multimodal MRI scans of the brain. The proposed radiomics approaches achieved an accuracy score of 80% on the limited test sets.