Human-level recognition of blast cells in acute myeloid leukaemia with convolutional neural networks

Microscopic examination and classification of blood cells is an important cornerstone of haematological diagnostics^1–3. Specifically, morphological evaluation of leukocytes from peripheral blood or bone marrow samples is one of the initial steps in the diagnosis of haematopoietic malignancies such as acute myeloid leukaemia (AML)^4–6. In particular, the common French–American–British (FAB) classification of AMLs strongy relies on cytomorphology⁷. Having been part of routine work-up of haematological diagnosis since the nineteenth century, cytomorphological examination of leukocytes has so far defied automation and is usually performed by trained human experts. Cytomorphological classification is thus tedious and time-consuming to produce, suffers from considerable intra- and inter-observer variation^8–10 that is difficult to account for, and is hard to deliver in situations where trained experts are lacking. Furthermore, it is difficult to reliably correlate with the results of other, intrinsically more quantitative diagnostic modalities such as immunophenotyping or molecular genetics⁴. An automated approach would allow for consistent classification of cytomorphologies, removing the intrinsically subjective human element of the process. Reliable, automated differentiation of cell morphology and recognition of malignant cells is also a key prerequisite to allow screening for haematological neoplasms, potentially enabling their earlier detection and treatment.

As cytomorphological examination is based on evaluating microscopic cell images, it can be formulated as an image classification task. Deep convolutional neural networks (CNNs) have proven very successful in the field of natural image classification^11–13. Recently, CNNs have been applied to various medical imaging tasks, including skin cancer recognition¹⁴, evaluation of retinal disorders¹⁵ and the analysis of histological sections^16,17, for example through mitosis detection¹⁸, region of interest detection and analysis¹⁹ or tissue type segmentation²⁰. This motivated us to apply CNNs to cytomorphological classification of blood cells, in particular those relevant in AML.

Previous work on leukocyte classification has mainly focused on feature extraction from cytological images^21,22. In that context, lymphoblastic leukaemias, where the cytomorphology is less diverse than in the myeloid case, have received more attention^23,24. Providing a sufficient number of labelled images for deep learning methods to work has proven challenging in medical image analysis due to restrictions on access to and the expense of expert time...