Machine learning in biomedical image analysis: predicting tumour growth and metastatic colonisation

In the era of artificial intelligence (AI), cutting edge techniques are being developed to tackle a variety of biomedical challenges. Moreover, it has long been recognised that automated image processing is more accurate and time-efficient than manual analysis. Indeed, the algorithms of machine learning supersede the human capability to extract relationships between different factors and allow to capture patterns invisible to the naked eye. The fast-growing field of computer-aided diagnosis encompasses applications of machine learning algorithms to biomedical data that provides rapid and reliable predictions and facilitates the decision-making processes in radiology and pathology. The vast body of this research addresses oncology. Years of clinical experience have proven that the key to successful treatment of cancer is early detection, diagnosis and accurate prediction of the disease progression. Medical imaging and histopathology are essential components of the cancer diagnostic procedure. It is now recognised that automated image processing augmented by AI can significantly advance screening and prognostic tools and therefore lead to a reduction in cancer mortality rates.

With this backdrop, my PhD research focussed on modelling tumour growth and predicting its progression by analysing image data at the cellular, tissue and organ levels. In a multidisciplinary research team, I investigated the key mechanisms of breast cancer growth and its ensuing metastatic colonisation in the liver by evaluating the invasion patterns and distributions of the cell populations. Our discovery of the characteristic cancer colonisation patterns in the liver required much more systematic scrutiny of the individual cancer cell behaviour under conditions mimicking the microenvironment of the liver being metastasised. My colleagues and I realised this biomimetic system using breast cancer cells cultured on a substrate with precisely controlled parameters, such as stiffness and material composition. Single-cell data analysis coupled with state-of-the-art machine learning methods was performed to unveil microenvironment-induced cancer cell plasticity through alterations of their molecular and morphological profiles. These findings have been reported in two publications.

In the first paper, we present a 3D model mimicking the process of liver colonisation and the formation of micrometastases by highly aggressive triple negative breast cancer (TNBC). My co-authors and I used a novel tissue engineering methodology to fabricate a whole-organ scaffold of the liver by removing native cells while preserving the organspecific architecture and biochemical composition of the extracellular matrix (ECM). The decellularised liver matrix was distinctly compartmentalised into parenchymal (soft and porous) and stromal (stiff and dense) sections. TNBC cells were seeded onto the matrix and cultured for a period of four weeks. Image analysis of the histological sections was performed to quantify cancer cell density, morphology and spatial distribution.

We observed profound differences in the cell growth and colonisation patterns between the parenchymal and stromal compartments. The chemotherapeutic drug resistance of TNBC cells in our tissue engineered constructs appeared greater in comparison to traditional monolayer cell cultures. Additionally, our engineered tumours were capable of inducing an angiogenic switch in the in ova chick embryo chorioallantoic membrane model. A sophisticated deep learning methodology was developed to accurately segment the intricate network of blood vessels and examine the angiogenic potential of our constructs.

In the second paper, the effects of matrix physical properties on the behaviour of TNBC cells were investigated. My research team partners and I focussed on exploring how the stiffness of the underlying substrate promotes morphological and phenotypic plasticity in TNBC cells. This plasticity is known to facilitate cancer cells in dissemination and metastatic colonisation. Using image-based cell profiling followed by comprehensive single-cell data analysis, we evaluated the characteristics of the cells cultured on the substrates of five different elastic moduli. This allowed me to link stiffness-induced changes in the cell morphology and structure to the expression of biomarkers associated with epithelial-mesenchymal transition (EMT). We demonstrated that cell growth and adaptation patterns change dramatically with substrate stiffness with a propensity for formation of multicellular clusters on the highest tested stiffness.

Aiming to further demonstrate the potential of AI for cancer research, my collaborators and I examined the application of deep learning techniques to tumour detection. Computer-aided detection (CAD) systems are suitable for analysis of medical images generated by various techniques including magnetic resonance imaging (MRI), X-ray or endoscopic imaging modalities. For instance, the key diagnostic procedure for bladder cancer is cystoscopy during which a clinician examines the inner lining of the bladder by inserting an endoscope through the urethra. In the third paper, I present a CAD system developed to assist in detection and diagnosis of urothelial carcinoma. Images acquired during cystoscopy examinations were interpreted by an expert and the boundaries of suspicious lesions were carefully annotated. The algorithm, based on a deep learning approach, is designed to identify the boundaries of papillary tumours in cystoscopic images by extracting highly complex features and computing the likelihood that a pixel belongs to a malignant lesion. The resulting segmentation masks provide guidance in the interpretation of cystoscopic images, and as such represent a successful outcome of this study.