Development of cellular and tissue segmentation algorithms within the digital pathology platform QuPath for accurate assessment of mIF across a breadth of solid tumours.

Professor John Le Quesne, Dr Peter Bankhead, Professor Joanne Edwards

Recent years have seen a revolution in our ability to generate high-resolution microscopic images of tumour tissue to reveal gene activity in single cells. These methods have the potential to answer a huge range of questions with relevance to basic cancer biology as well as improving outcomes for patients. In particular, advancements in multiplex immunofluorescence staining (mIF) using such technologies as POLARIS and CODEX, and in the physics of multiplex image microscopy, now make it possible to acquire microscopic images with numerous channels each of which quantifies a single gene product at RNA or protein level. The challenge is to convert these images into quantitative single-cell data for subsequent analysis using methods such as spatial statistics and artificial intelligence. The quality of the data depends upon our ability to segment images into regions (eg tumour epithelium/stroma/ necrosis/normal) and into single cell compartments (nuclei/cytoplasm/membranes). There is as yet no widely accepted transferable method to achieve this accurately. Within the CRUK Scotland Major Centre we have all the resources, physical and intellectual, required to generate a world-class segmentation algorithm which will then be used to underpin a huge range of projects within the centre. We are extremely active in the collation of patient image cohorts for biological discovery, with H&E and mIF images from many thousands of patients with common malignancies (Le Quesne/Edwards groups in Glasgow). In parallel, the Bankhead group in Edinburgh are creators of the open-source QuPath digital image analysis platform and leading experts in the development of image analytical algorithms with a particular interest in the segmentation problem. QuPath provides extensive tools to visualize and query multiplexed data, and is now one of the most widely-used platforms for digital pathology analysis (>200,000 downloads, >1,100 citations on Scopus). Deep learning artificial intelligence is being built into the QuPath code, and the successful applicant will apply AI methods to real-world tumour images to achieve optimal segmentation from image information. The project will initially focus on lung and colon cancer images, with the ambition of generating a generally applicable pan-cancer segmentation algorithm. The focus is on development of the segmentation method, but numerous biological benchmark phenomena will be quantified in the course of study, with the potential for further high-impact publications. In particular, we would bring the segmentation algorithm to bear upon active research questions around the immune microenvironment and tumour architecture.


• generate a pan-cancer QuPath-based cell segmentation method (nucleus, cytoplasm, membrane)

• generate a pan-cancer QuPath-based tissue segmentation method (malignant epithelium, stroma, necrosis, normal tissues)

• validate the method across multiple tumour types and against various ground truths (eg human assessment, clinical outcomes)

Lab websites