Deciphering Blood Cells - Method for Blood Cell Analysis using Microscopic Images

Abstract

Blood cell analysis of RBC and WBC from smear images is crucial for identifying disorders caused by abnormal blood cells, such as leukemia, anemia, frequent infections, and other diseases, including cancer. Pathologists find it challenging to distinguish between abnormal and normal cells in smear images. Manually counting each element is tedious and prone to human errors. Image processing techniques are used to enhance the smear image quality before segmentation. Identifying and segmenting smear images remains a challenge for pathologists. In this research, we propose an approach for identifying and segmenting smear images, followed by classifying normal and abnormal cells. This work implements Mask CNN for segmentation and two different algorithms: EfficientNet-B3 and DenseNet-121 for classification. The Mask CNN model's results for identification and segmentation are evaluated using mAP and IOU. For RBC, the best mAP value is 0.52 and IOU is 0.45. For WBC, the best mAP value is 0.56 and IOU is 0.52. Classification results outperform previous studies, with experiments on six different classes based on RBC and WBC cell normality and abnormality. The average overall accuracy for EfficientNet-B3 is 95% and DenseNet-121 is 97%, respectively.