NPCA-wbcNet: advanced neuron-position correlation attention network for white blood cell classification via optical microscopic imaging in leukemia diagnosis

Abstract

Optical microscopic imaging is crucial for identifying and enumerating different types of white blood cells (WBCs), which are essential for diagnosing hematological disorders like leukemia. However, the subtle morphological differences among various WBC types pose challenges to accurate classification, even for experienced experts. Deep learning has emerged as a powerful tool to enhance optical microscopic imaging in WBC classification. Therefore, this study proposes a novel convolutional neural network with neuron-position correlation attention (NPCA) named NPCA-wbcNet to overcome this challenge of WBC classification. Specifically, the NPCA could effectively generate 3D weights to infer the discrimination and importance of each neuron and establish interactive relationships between different locations and channels, thereby highlighting the location and shape of the target category in the image and enhancing the identification of fine-grained features. Additionally, we replace some regular convolutions with partial convolutions, which makes the model more lightweight and achieves better results in extracting spatial features. We evaluate the performance of our model on DrlWBCs, a new dataset of 24,713 blood smear images with five WBC subtypes. The model achieved an accuracy of 98.46 % for WBC classification, and the acute lymphoblastic leukemia (ALL) diagnostic accuracy on the ALL-IDB2 dataset was 98.08 %. Overall, our method surpasses other state-of-the-art methods in terms of accuracy, precision, and F1-score performance metrics.