SegMWB: A lightweight deep learning framework for microscopic image classification

White Blood Cells (WBCs) play a significant role in assessing an individual's health. An automated WBC classification system is desirable to diagnose various hematological malignancies at an early stage. This work proposes a new deep learning-based framework, i.e., SegMWB, for classifying diverse microscopic images of white blood cells. The proposed framework consists of pre-processing, segmentation and classification phases. The pre-processing phase performs transformation, normalization and augmentation. A customized nucleus segmentation algorithm is designed to extract more discriminative features from cell images. The classification phase includes the proposed SegMWB-Net, which integrates the block-wise pattern of convolutional layers and allows the framework to extract more meaningful information from the multiple regions, making the model more efficient for capturing the local features of cell images and simultaneously reducing the complexity of model. A batch normalization layer is employed to reduce the likelihood of overfitting and to help the network converge faster. The proposed framework was evaluated on three publicly available microscopic WBC datasets acquired under different imaging conditions, including variations in resolution, staining and illumination. SegMWB achieved competitive classification performance with accuracies of 96.54%, 98.37% and 98.70% on the Raabin-WBC, Peripheral Blood Cell and LISC datasets, respectively. The results indicate that the proposed segmentation-assisted lightweight deep learning framework can improve WBC classification while maintaining lower computational complexity than several transfer-learning-based models. However, further validation on independent clinical datasets is required before considering practical clinical deployment.