In the realm of haematology, the transition from manual microscopic examination to automated cell morphology analysers as the gold standard remains an unachieved milestone. A pivotal impediment to this advancement is the inadequacy in the volume and fidelity of labelled cellular samples essential for the training of artificial intelligence (AI)-driven cell classification models. The process of deriving labeled cells from digital microscopic imagery necessitates meticulous human curation, which, given the extensive quantity of cells required, becomes onerous and fraught with quality control challenges.
This manuscript introduces an innovative methodology designed to significantly enhance the efficiency of the labelling process. Initially, we employed some techniques to mitigate the confounding effects of erythrocyte populations on leukocyte identification, thereby expediting subsequent cellular sorting procedures. Subsequently, leveraging state-of-the-art cell sorting technology, we executed label-free segregation of a targeted leukocyte subset, exemplified by monocytes, culminating in the acquisition of a highly purified monocyte suspension. We adhered to established slide preparation and staining protocols, such as the Wright--Giemsa staining method, to fabricate blood smears that preserve cellular morphology with minimal alteration. In the final phase, human annotators perform batch labelling of monocytes through the mediation of digital microscopic imagery of the smears. Consequently, the resultant digital microscopic images predominantly feature monocytes, with a negligible presence of other leukocyte classes.
By delineating non-monocytic cells with the annotation "not monocytes," annotators can efficiently designate the remaining cells as monocytes, thereby achieving batch labelling of the specified leukocyte class.
This approach not only augments the efficacy of manual labelling but also diminishes the laboriousness associated with the task, with the anticipation of concurrently elevating the calibre of labelling accuracy.
