Parallel Wavelet Network Achieves 94.4 Percent Abnormal White Blood Cell Accuracy

A newly developed parallel classification network termed PV-WTConv achieved 94.4 percent accuracy in classifying abnormal white blood cells on the AWBC-LZWH image dataset, according to an original research study published in Frontiers in Medicine. The architecture, which integrates wavelet transform convolution, demonstrated performance gains over the evaluated reference models.

When evaluated against baseline deep learning frameworks on the AWBC-LZWH test set, PV-WTConv substantially outperformed a standard convolutional baseline, which registered an accuracy of 56.9 percent. This represented an absolute difference of 37.5 percentage points. The proposed parallel network also exceeded the performance of the established ResNet50 architecture, which achieved 91.4 percent accuracy on the same benchmark, reflecting an absolute improvement of 3.0 percentage points.

The authors contextualized their model development within cell detection research, emphasizing that identifying abnormal white blood cells is critical for the early clinical diagnosis of malignant haematological diseases, including leukaemia and myelodysplastic syndromes. In clinical practice, evaluating white blood cell abnormalities additionally serves as an essential indicator for monitoring immunodeficiencies, assessing infectious diseases, and tracking the therapeutic efficacy of chemotherapy regimens.

The researchers highlighted that abnormal white blood cells frequently display blurred boundaries distinguishing them from normal cells or cells exhibiting natural physiological variations. Furthermore, different abnormal cell types, including leukaemic cells and hyperplastic cells, lack uniform and fixed morphological criteria, while variations in cell staining techniques and imaging complexity introduce significant visual interference.

To train and evaluate the parallel network, researchers from Guangxi Science and Technology University and Liuzhou Workers’ Hospital relied primarily on the AWBC-LZWH dataset of abnormal white blood cell images. This primary image collection was provided by the Department of Clinical Laboratory Medicine at Liuzhou Workers’ Hospital and supported the study’s main training, validation, and testing phases. Additionally, the investigators conducted supplementary experiments utilizing the PBC_dataset, a public collection of peripheral blood cell images assembled and released by the Hospital Clínic de Barcelona.

Although PV-WTConv demonstrated strong classification capabilities across the evaluated peripheral blood and abnormal white blood cell image datasets, the study authors explicitly noted that the network’s real-world clinical diagnostic value remains unestablished. They concluded that its clinical diagnostic value still requires further validation through prospective, multicentre studies.

Subject of Research: Computer-aided classification of abnormal white blood cells using wavelet transform convolution

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

Article References: Fan, X., Song, X., Sun, B., Yao, Z., & Meng, J. (2026). PV-WTConv: a parallel network for abnormal white blood cell classification based on wavelet transform convolution. Frontiers in Medicine, 13, Article 1947482. https://doi.org/10.3389/fmed.2026.1947482

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1947482

Keywords: abnormal white blood cells, wavelet transform convolution, PV-WTConv, white blood cell classification, AWBC-LZWH dataset, PBC dataset, ResNet50, computer-aided diagnosis, leukaemia detection, myelodysplastic syndromes, peripheral blood cell images, microscopic cell imaging, morphological cell criteria, haematological diseases, parallel neural networks, chemotherapy efficacy monitoring, immunodeficiency assessment