Semantic Segmentation of Leukocytes: Aiding Morphological Analysis by Attention-Guided U-Net Leveraging Residual Convolutional Blocks

Evaluating the morphology of peripheral leukocytes is vital for research in medical fields, encompassing the treatment and diagnosis of illnesses such as immunodeficiency and blood cancer. Conventional detection methods are prone to interference, profuse labor, and increased time consumption. Additionally, efficient segmentation becomes quite challenging due to variations in cell types, staining procedures, and intercell adhesion. This research aims to utilize the Raabin-WBC dataset and compare the multiclass segmentation performance, i.e., background, cytoplasm, and nucleus, of three convolutional neural network architectures: U-Net, Attention U-Net, and Residual Attention U-Net. The IoU, Precision, Recall, and Dice Coefficient for Residual Attention U-Net were found to be 91.9%, 95.4%, 96.1%, and 95.7%, respectively. Besides detecting false negatives, the Residual Attention U-Net significantly outperformed the other two models.