Artificial Intelligence assisted mapping of prostate cancer progression in patient biopsies with novel tissue labelling biomarkers - beyond Gleason Sc

点击翻译按钮获取中文摘要

The prostate gland is the site of the most common cancer in men, with 1.2 million new cases diagnosed annually worldwide. The main diagnostic pathway requires MRI followed by biopsy to obtain needle core tissue. Histopathology analysis of the samples is used to diagnose the stage of cancer based on the long-established Gleason score that classifies the morphological changes of the tissue, with higher scores associated with more aggressive disease and worse prognosis. The Gleason scale was established in the 1960's and updated in 2005 and 2014 by international conferences of experts. Refinements were centred around specific types of formation, such as cribriform glands, glomeruloid glands, and mucinous carcinoma, associated with aggressive disease. Nevertheless, Gleason scoring remains a pathologist-led diagnostic modality, based on experience. Over the past decade, artificial intelligent tools have demonstrated the ability to assist prostate cancer diagnosis in biopsies by employing image processing and pattern recognition technology. Standardisation, reproducibility, and provenance, however, are still not adequately addressed. Moreover, omic data and spatial biomarker readouts have not been integrated into the AI-assisted diagnostic process. We propose to develop AI and Machine Learning algorithms to integrate multiple streams of phenotypic and omic data into a novel in silico framework that characterises the prostate biopsy tissue for cancer stage and progression. Tissue phenotyping will include the clinical standard of hematoxylin and eosin (H&E) staining along with novel tissue paints, such as DRAQ5, which selectively binds to nuclei an essential feature of cancer stratification. Additionally, spatial transcriptomics Nanostring technology data will be used to identify features of the stroma. Tissue matching omic data from the European Stem Cell Institute will also be used to formulate a novel multi-dimensional space approach, suitable for AI and machine learning tools to be developed and tested. We will employ both commercially available packages, such as Tensorflow, and Keras in Python, U-Net convolutional networks, as well as in-house developed systems to facilitate novel multi-physics data integration. The discovery element of the work will be supported by the fundamental phenotyping of the tumour microenvironment that has been the main focus of our research group over the past two decades