Using machine-learning approaches to identify blood cancer types

Myeloproliferative Neoplasms (MPNs) are a group of blood cancers that occur when stem cells in the bone marrow develop mutations that lead to over-production of blood cells – either red blood cells in Polycythaemia Vera (PV), or platelets in Essential Thrombocythaemia (ET). This carries an increased risk of developing blood clots, such as in the legs, lungs, heart attacks or strokes.

In myelofibrosis, the most severe of the MPNs, destructive scarring (‘fibrosis’) of the bone marrow develops, leading to failure of the marrow to produce blood cells and severe symptoms. Patients with all MPNs are at higher risk of developing leukaemia, especially patients with myelofibrosis when this develops in over 1 in every 10 patients.

Unfortunately, we do not yet have any drug treatments that can cure these conditions. Treatments for ET and PV aim to control the blood counts and reduce the risk of blood clots. For myelofibrosis, targeted therapies such as ruxolitinib, a JAK inhibitor, can effectively control symptoms, but this does not alter the natural history of the disease and survival remains less than 5-10 years following diagnosis.

In the vast majority of cases, mutations are found in one of 3 genes – JAK2, CALR or MPL. Screening for these is important in MPN diagnosis, however distinguishing between the MPN subtypes requires a careful examination of blood counts and the morphological features of a bone marrow biopsy.

Unfortunately, assessment of the bone marrow is highly subjective, reliant on qualitative observations and there is great variability, even when it is done by expert haematopathologists. In particular, it is very hard to reliably distinguish between a mutation-negative MPN and a ‘reactive’ (non-cancer) bone marrow.

A more accurate method for diagnosis is very much needed, to enable selection of the most appropriate treatment strategy for patients and to determine treatment targets. Megakaryocyte cells or  ‘megas’ – the large, bone marrow cells that produce blood platelets – are very abnormal in all the MPNs and thought to play a key role in the disease pathology. Interestingly, although the gene mutations underlying all 3 MPNs lead to an over production of megas, subtle differences in the appearance and location of these cells within the bone marrow occur in the different MPN subtypes.

To try to improve MPN classification, a team lead by Jens Rittscher (Department of Engineering) and Daniel Royston (Radcliffe Department of Medicine), developed an AI approach to screen and classify MPN cases based on features of the mega cells, discovering new features in their cell size, clustering and internal complexity. Their machine learning approach revealed that there are clear differences between MPN subtypes – the platform was able to more accurately classify patients by assessing subtle morphological differences in the biopsies that could not have been identified by the naked eye.

These findings have been published in Blood Advances. Dr Beth Psaila, a clinician scientist at the MRC Weatherall Institute of Molecular Medicine and a haematology consultant specialising in MPNs said:

“It has long been recognised that a multitude of subtle differences in megakaryocyte morphology can distinguish between the MPN subtypes. However, this means that assessment of bone marrow biopsies is poorly reproducible, sometimes leading to diagnostic uncertainty and inappropriate treatment plans for patients.

“The approach developed here is really exciting for the field, as it is now possible to perform deep phenotyping of megakaryocytes and more accurate disease classification using simple H&E slides which are routinely prepared in all diagnostic facilities. This will be incredibly useful both for research aimed at better understanding the role of megakaryocytes in blood cancers as well as improving diagnosis and treatment pathways for our patients.”

The team hopes that in the future, this work can be combined with other histological assessments to optimise the clinical application of AI approaches, and create a more comprehensive quantitative description of the bone-marrow microenvironment and its cancers.

About the researchers and the study

This work was funded by the NIHR Oxford Biomedical Research Centre and is the result of collaboration between Korsuk Sirinukunwattana (Department of Engineering), Alan Aberdeen (Ground Truth Labs Ltd.), Helen Theissen (Department of Engineering), Jens Rittscher (Department of Engineering) and Daniel Royston (Radcliffe Department of Medicine [NDCLS]).

Jens Rittscher is a Principle Investigator whose research aim is to enhance our understanding of complex biological processes through the analysis of image data that has been acquired at the microscopic scale. Jens develops algorithms and methods that enable the quantification of a broad range of phenotypical alterations, the precise localisation of signalling events, and the ability to correlate such events in the context of the biological specimen.

Korsuk Sirinukunwattana is a postdoctoral research assistant in Rittscher’s group specialised in medical image analysis and computational pathology. His main research interest is the association between tissue morphology and molecular/genetic subtypes in various diseases.

Alan Aberdeen leads Oxford spinout Ground Truth Labs, a company supporting digital pathology research through on-demand analysis, biomarker discovery, and high-quality cohorts.

Helen Theissen is a doctoral research student in Rittscher’s group. Her research focuses on computational methods to characterise cellular subtypes and quantify the bone marrow microenvironment in MPNs.

Daniel Royston is a joint academic & consultant Haematopathologist at Oxford University Hospitals NHS Foundation Trust / Radcliffe Department of Medicine.