Research focus

Machine learning for digital pathology

Our lab applies technological innovations in machine learning and computer vision to large-scale applications in digital pathlogy. These applications enable us to map the spatial distribution of heterogeneous cell types within tumour section images at single-cell resolution.

Cancer as evolving ecosystems

Our scientific focus is on understanding tumours as evolving ecosystems. This is an emerging concept, that tumours are complex, evolving ecosystems with dynamic crosstalk among multiple cell types. Such crosstalk could ultimately dictate the evolutionary trajectory of cancer, thereby influencing cancer progression and therapeutic response.

Just like in ecology where spatial organisation of animals, their predators and habitats is central for understanding the ecosystem, it is becoming increasingly evident that we need to use a similar, spatially explicit approach to study cancer. Thus far, our work has changed the way we think of the microenvironment and its critical roles in different cancer types.

Image-Omics data integration

Understanding how genetically diverse cancer cells adapt to their microenvironment is key to appreciating how they develop and evolve. We combine image-based cell spatial mapping with bioinformatics and ecological statistics to studies how different cancers evolve and spread amidst healthy tissue at high spatial resolution.

Therefore, we want to understand why cancer is so difficult to treat, and use ecololgical principles to direct more effective therapeutic interventions - akin to draining the swamps to get rid of mosquitoes as one way of eradicating malaria. By building a highly interdisciplinary programme, our goals are to deliver scientific advances and clinical innovations by dissecting the spatial heterogeneity of tumour microenvironment, to foster new developments in machine learning for pathology, and to formulate objective methodologies for directing cancer therapeutic strategies.


Dissecting cancer proliferation heterogeneity with deep learning

Our development of a Simultaneous Detection and Cell Segmentation (DeepSDCS) network to perform cell segmentation and detection in Ki67-stained immunohistochemistry images was accepted for MIDL 2018 (link). This framework combines VGG16 with hypercolumn descriptors to form the vector of activation from specific layers to capture features at different granularity. Features from these layers were propagated using a stochastic gradient descent optimizer to yield accurate detection of cell nuclei and the final cell segmentations.

Spatial statistics to understand breast cancer recurrence

Our recent study on 1178 breast cancer patients underscored the importance of examining spatial heterogeneity of the tumour. We studied how immune cells are spatially arranged within the tumours, and detected the so-called immune hotspots, which are tumour regions that contain spatial clustering of immune cells. This uses a spatial statistical method called Getis-Ord Hotspot analysis, which is commonly used for detecting crime hotspots in cities. High amount of immune hotspots, but not the amount of immune cells, correlates with high probability of cancer recurrence. This study provides a new way to predict patient prognosis, and open the door to new therapeutic opportunities using immunotherapy for breast cancers.


Centre for Evolution and Cancer & Division for Molecular Pathology,
The Institute of Cancer Research, London
The Royal Marsden Hospital

Join us!

Summer student

An opportunity to be part of the team that uses artificial intelligence / deep learning to better study breast cancer in one of the largest clinical trials in the UK. This position will assist with digitalising pathological samples and quality control for about 2 months, contributing to an ambitious study using a total of 9,000 digital pathology samples to understand cancer treatment resistance, a major cause of cancer death. You will need to be over 18 years old. Email yinyin.yuan (a) with your CV.

Postdoc positions

Image processing This is an exciting project collaborating with Prof. Janet Shipley to develop novel deep learning approaches for rhabdomyosarcomas. Rhabdomyosarcomas are rare cancers but a major cause of death from cancer in children. Please enquire by emailing (yinyin.yuan (a) with your CV. Apply here!

Machine learning Studying lung cancer within the TRACERx lung cancer program, in collaboration with Prof. Charles Swanton at the Francis Crick Institute Please enquire by emailing (yinyin.yuan (a) with your CV. Apply here!

Selected publications

Microenvironmental niche divergence shapes BRCA1-dysregulated ovarian cancer morphological plasticity
Heindl A, Khan AM, Rodrigues DN, Eason K, Sadanandam A, Orbegoso C, Punta M, Sottoriva A, Lise S, Banerjee S, Yuan Y
Nature Communications (in press)

Understanding morphological diversity in the spatial context of ecological environment was fundamental to the discovery of Darwinian evolution. We provided empirical evidence supporting the selective advantage of a cancer cell subpopulation with morphological diversification in locally immunosuppressive microenvironment, signifying morphological diversification as an ecological process of cancer with profound clinical implications.

Relevance of spatial heterogeneity of immune infiltration for predicting risk of recurrence after endocrine therapy of ER+ breast cancer.
Heindl A, Sestak I, Naidoo R, Cuzick J, Dowsett M, Yuan Y*
JNCI (2018) [Additional ppt]

We provide a missing link between tumour immunity and disease outcome in ER+ disease by examining tumour spatial architecture. The association between immune spatial scores and late recurrence suggests a lasting memory of protumour immunity that may impact disease progression and evolution of endocrine treatment resistance, which may be exploited for therapeutic advances.

Microenvironmental heterogeneity parallels breast cancer progression: A histology-genomics integration analysis
Natrajan R, Sailem H, Mardakheh FM, Arias MG, Dowsett M, Bakal C, Yuan Y*
PLoS Medicine (2016) [Sweave; R package beta version]

We propose a clinically relevant role of tumour microenvironmental diversity for advanced breast tumours and highlight that ecological statistics can be translated into medical advances for identifying new biomarkers and for understanding the synergistic interplay between the microenvironment and cancer genomics.