As part of our EPSRC funding, we have been awarded £400,000 to be allocated to pump-priming projects that will accelerate the integration of SHP, DP and AI. These are to be broadly aligned with our work packages but not limited to them. Applications are welcome from anyone who is eligible for EPSRC funding and will be based primarily on scientific excellence. However, preference will be given to those who are actively engaged with the Network and have attended the Sandpit session where the Network Challenges will be discussed. We envisage funding 8-10 projects and we are limiting each project to a maximum of £50k. Please note that, in line with EPSRC policy, we can only award 80% of the cost of the project. Applications will be open immediately after each sandpit event and details of how to apply will be found on the applications page.
Below are some of the projects successfully awarded funding to date.
Breast Cancer Receptor Status Prediction through Vibrational Spectroscopy;
Fayyaz Minhas, University of Warwick
Peter Gardner, University of Manchester; Valerie Speirs, University of Aberdeen; Alex Henderson, University of Manchester; Nicholas Stone, University of Exeter / Royal Devon and Exeter Hospital
MSInRed: Digital Pathology and AI based Association and Integration of Histological and Molecular Fingerprints of Microsatellite Instability in Colorectal Cancer
Nasir Rajpoot, University of Warwick
David Snead, University Hospitals Coventry & Warwickshire NHS Trust / Department of Pathology; Klaus Gerwert, Ruhr University Bochum, Department of Biophysics and Centre for PROtein DIagnostics (PRODI); Axel Mosig, Ruhr University Bochum, Department of Biophysics and Centre for PROtein DIagnostics (PRODI); Frederik Großerüschkamp, Ruhr University Bochum, Department of Biophysics and Centre for PROtein DIagnostics (PRODI)
Colorectal cancer, a prevalent form of cancer impacting the large intestine, has garnered attention in recent medical research for tests detecting microsatellite instability (MSI), a marker influencing treatment decisions, particularly for immunotherapy efficacy. Our study aims to identify MSI using two image modalities: conventional pathology images observed under a microscope and infrared images capturing molecular details unattainable in standard imaging. Through the development of deep networks for classification, we analysed images from 760 colorectal cancer patients to discern patterns indicative of MSI. Evaluation of whole slide images from Bochum University revealed a significant correlation between observed patterns and MSI presence.
Vibrational spectroscopy based prediction of spatial transcriptomic profiles
Fayyaz Minhas, University of Warwick
Nicholas Stone, University of Exeter / Royal Devon and Exeter Hospital; Jayakrupakar Nallala, University of Exeter; Mudassar Iqbal, University of Manchester, Division of Informatics, Imaging and Data Sciences; Syed Murtuza Baker, University of Manchester, Bioinformatics Core Facility, Faculty of Biology, Medicine & Health; Kevin Couper, University of Manchester, Division of Immunology, Immunity to Infection and Respiratory Medicine; Martin Fergie, University of Manchester, Division of Informatics, Imaging and Data Sciences; Federico Roncaroli, University of Manchester, Division of Neuroscience and Experimental Psychology
This project aimed to develop a deeper understanding of tissue biology in diseases such as brain cancer by combining advanced technologies, such as machine learning, spatial transcriptomics, and spectroscopy. These methods allow scientists to study tissues at a detailed level, revealing the interactions between genes and proteins within specific locations in the body. While obtaining spectral data, we made significant progress by developing a powerful tool named Ouroboros. This tool predicts protein expressions from standard tissue images and can generate synthetic tissue images from protein data allowing us to cross-link these modalities. Ouroboros represents a significant advancement, offering a new way to study tissue samples, potentially leading to better diagnostic methods which can be used with spectral data. Additionally, all related (anonymised) data is compiled and available online for future research. This project illustrates how combining different scientific disciplines can lead to innovative solutions in medical research and patient care.
Spectral Phasor-based Automated Raman Histopathology and Tissue Analysis (SPARTAN)
Karen Faulds, The University of Strathclyde
Valerie Speirs, University of Aberdeen; Rasha Abu-Eid, University of Aberdeen; Katie Hanna, University of Aberdeen; William Tipping, University of Strathclyde
This project will develop histopathology of breast cancer tissue biopsies using stimulated Raman scattering (SRS) microscopy. SRS microscopy combines label-free imaging with high spatial visualisation which is ideally suited to imaging biological samples. The project will assess the feasibility of using spectral phasor analysis (SPA) for automatic segmentation of hyperspectral SRS imaging across biopsy samples. Spectral phasor analysis is a label-free processing technique that enables the segmentation of biological samples based directly on the SRS spectrum. The pixels are clustered into regions of similar spectral profiles (and hence, similar biochemical composition). This project will develop automated segmentation of the spectral phasor plot derived from patient biopsies for the identification of malignant and healthy regions of the tissue.
Integrating infrared spectroscopy and digital pathology: an explainable AI multi-modal framework for breast cancer diagnosis
Annalisa Occhipinti, Teesside University
Claudio Angione, Teesside University; Pietro Liò, University of Cambridge; Peter Gardner, University of Manchester; Alex Henderson, University of Manchester; Abhik Mukherjee, NHS; Jessica Eyssautier, NHS
Cancer diagnosis currently relies mostly on imaging and tissue analysis, which can make the process prone to errors due to the subjective interpretation of the pathologists. Our project proposes a combined approach (based on multi-modal AI) that integrates histopathological images (detailed pictures of tissue samples) with infrared spectroscopy (a technology that looks at the chemical properties of tissues) to provide comprehensive insights into tumour progression. The proposed approach has the potential to (i) enhance the accuracy and reliability of cancer diagnosis by leveraging the strengths of both data modalities and (ii) complement the morphological details obtained from histopathological images with molecular spectroscopy information (i.e., information about the parts that make up the tissue).
Distinguishing Chondrosarcoma Subtypes with Raman Spectroscopy vs Digital Histopathology: A Pathway towards Clinical AI Studies
Tapabrata Rohan Chakraborty, University College London
Adrienne M Flanagan, University College London, Cancer Institute; Geraint M. H. Thomas, University College London
Chondrosarcoma is a type of cancer of connective tissues and there a multiple subtypes, some of which are particularly challenging to distinguish between. Determining these subtypes is critical in order for patients to receive the appropriate treatment for them. Raman spectroscopy (RS) is a technology that uses light which can detect subtle biochemical changes. This may be useful to detect chondrosarcoma subtypes, but this technology results in more data than a human can analyse. Artificial Intelligence (AI) could help revolutionise the diagnosis of cancers as it can analyse vast amounts of medical data. In our research we will apply AI to two types of data, RS data vs standard digital pathology images used in the clinic for grading/sub-typing based on visual manifestation of clinical features. We will use both methods to compare their effectiveness. If neither is good enough on its own, we will try combining them. We will also look at how long each method takes, how reliable they are, and how well they can be explained and understood by medical experts.
Spectraspace: Spectral-Spatial Attention Models for Molecular Spectroscopy and Digital Histopathology
Duygu Sarikaya, University of Leeds
Marc de Kamps, University of Leeds; Pietro Lio, University of Cambridge; Annalisa Occhipinti, Teesside University; Claudio Angione, Teesside University; Aidan D. Meade, Technological University Dublin; Anna Nowakowska, Jagiellonian University in Kraków; Axel Mosig, Ruhr University Bochum
Multi-modal AI models are proposed in the literature to distinguish diseased from non-diseased tissue using both hyperspectral imaging and digital pathology, as the spectral and spatial data are complementary and might provide additional cues that the other modality is missing. These models learn richer representations which improves robustness and performance for the defined tasks. However, state-of-the-art AI models designed to work with specific bands of the spectrum (such as RGB in visible light) fail to efficiently make use of the wealth of information provided by hyperspectral imaging. While some of these continuous spectral bands carry more relevant information, others might be carrying less relevant or redundant information. Moreover, these models may have a strong inductive bias towards morphology learned from the spatial data rather than spectral features. Effectively using both spatial and spectral information remains a major challenge. We will use attention models to make use of the most relevant and important spectra as well as to use this information efficiently.