MCNet: Mask Cell of Multi Class Deep Network for Blood Cells Detection and Classification

Abstract

Physicians are likely to expend significant labor and time while manually calculating blood smears. Automatic computer-based methods for classifying acute lymphoblastic leukemia have trouble correctly lighting stained white blood cell microscopy images and accurately separating cells that touch or overlap. Additionally, incorporating machine learning techniques into medical services is very hard because doctors can deal with rough guesses as long as the results aren't too bad, but they can't use these calculations for actual medical care. Enabling a deep network to have knowledge of the accuracy of its own predictions is a fascinating and crucial issue. Most instance segmentation frameworks weigh the mask quality during the instance segmentation process based on classification confidence. Here, we consider the context of this problem and present Mask Cell of multi class deep network (MCNet) as a new network that has the module to learn about the quality of the predicted instance masks. Our proposal entails using faster R-CNN, such as segmentation on white blood cell microscope images, to accurately categorize acute lymphoblastic leukemia cases. This approach aims to enhance the efficiency and effectiveness of the diagnostic process. The suggested network block combines the instance feature with the matching anticipated mask to estimate the proposed mask IoU. In this work, we used the transfer learning approach to apply Mask R-CNN to segment white blood cells on a microscope image. To address the issue of poor lighting in stained white blood cell microscopy pictures, we included a contrast enhancement procedure in the image dataset. The comparative experiment applies YOLO v9 for classification and Mask R-CNN. The MCNet approach adjusts the discrepancy between the quality of the mask and its proposed detection, enhancing the effectiveness of instance segmentation. The final results for two datasets trained using PBC and BCCD are as follows: the accuracy of mAP@IoU0.50 for the PBC dataset is 95.70, while the accuracy for the BCCD dataset is 96.76, with recall and precision both coming in at 97.23 and 96.72 respectively.