CompTRNet: An U-Net Approach for White Blood Cell Segmentation Using Compound Loss Function and Transfer Learning with Pre-trained ResNet34 Network
Md. Jibon Mia, Md. Mossadek Touhid, Sunanda Das · 2022
Identification of White Blood Cells (WBCs) is a bold problem because of having different varieties of cells and the variability present in human perpheral blood smear samples. It is important and needed for detection of blood related diseases such as leukemia, malaria, anemia, syphilis, etc to help the hematologists. Old methods for this identification are weighty and imprecise. Therefore, In this paper, we propose an encoder-decoder based U-Net architecture with ResNet34 backbone for leukocyte image segmentation. Specially, Backbone ResNet designs a feature encoder to bring out multi-scale features and establishes skip connections on condensed convolutional layers. Here, we also proposed a compound loss function (Dice loss & Focal loss) that not only considers pixels, but it also accounts for precision loss and recall loss. The segmentation attainment is evaluated on test set. The proposed method was able to achieve ME, FPR, FNR and Dice-coefficient of 0.0131, 0.0116, 0.0317, and 0.9665 respectively. The extensive result analysis indicates that the suggested method outperforms numerous existing approaches in terms of WBC segmentation.