Leukaemia stem cells (LSCs) are a rare population among the bulk of leukaemia cells and are responsible for disease initiation, progression/relapse and insensitivity to therapies in numerous haematologic malignancies. Identification of LSCs and monitoring of their quantity before, during, and after treatments will provide a guidance for choosing a correct treatment and assessing therapy response and disease prognosis, but such a method is still lacking simply because there are no distinct morphological features recognisable for distinguishing LSCs from normal stem cell counterparts. Using artificial intelligence (AI) deep learning and polycythemia vera (PV) as a disease model (a type of human myeloproliferative neoplasms derived from a haematopoietic stem cell harbouring the JAK2V617F oncogene), we combine 19 convolutional neural networks as a whole to build AI models for analysing single-cell images, allowing for distinguishing between LSCs from JAK2V617F knock-in mice and normal stem counterparts from healthy mice with a high accuracy (> 99%). We prove the concept that LSCs possess unique morphological features compared to their normal stem cell counterparts, and AI, but not microscopic visualisation by pathologists, can extract and identify these features. In addition, we show that LSCs and other cell lineages in PV mice are also distinguishable by AI. Our study opens up a potential AI morphology field for identifying various primitive leukaemia cells, especially including LSCs, to help assess therapy responses and disease prognosis in the future.