PV-WTConv: a parallel network for abnormal white blood cell classification based on wavelet transform convolution

In the field of cell detection research, the detection of abnormal white blood cells is not only central to the early diagnosis of malignant haematological diseases such as leukaemia and myelodysplastic syndromes, but also serves as a key indicator for assessing infectious diseases, immunodeficiencies and the efficacy of chemotherapy. The use of computer-aided techniques to accurately classify white blood cells in microscopic images can assist clinicians in diagnosing and detecting various blood-related diseases more promptly and accurately. However, the accurate identification and classification of white blood cells remains challenging due to the blurred boundaries between abnormal white blood cells and normal white blood cells—or even those exhibiting physiological variations—the lack of uniform and fixed morphological criteria for different types of abnormal cells (such as leukaemic cells and hyperplastic cells), and interference from factors such as the complexity of cell imaging and staining techniques. Therefore, this paper proposes a parallel classification network for abnormal white blood cells based on wavelet transform convolution, designed to enhance the accuracy of recognising and classifying different categories of abnormal white blood cells. This study utilised the AWBC-LZWH dataset of abnormal white blood cell images, provided by the Department of Clinical Laboratory Medicine at Liuzhou Workers’ Hospital, for the main training, validation and testing phases, and conducted supplementary experiments using the PBC_dataset of peripheral blood cell images, which was collected and publicly released by the Hospital Clínic de Barcelona. PV-WTConv achieved an accuracy of 94.4% on AWBC-LZWH, compared with 56.9% for the convolutional baseline and 91.4% for ResNet50, corresponding to absolute differences of 37.5 and 3.0 percentage points, respectively. The results of this study indicate that PV-WTConv demonstrates good classification performance on the evaluated white blood cell image datasets; however, its clinical diagnostic value still requires further validation through prospective, multicentre studies.