Cell image instance segmentation based on PolarMask using weak labels

A PolarMask-based method for blood cell contour segmentation is proposed. The method is divided into two parts. One part is a weak label-based model pretraining method, which uses weak labels to train the model and obtain a pretraining weight. The training speed and accuracy of the segmentation model are accelerated. The other part is based on the PolarMask method to segment the white and red blood cells in blood cells and can obtain smoother cell contours. Thus, it improves the accuracy of blood cell segmentation. Our method can help medical personnel identify the number of cells and cell shape quickly, which reduces the workload for medical personnel. We improve PolarMask to make it more suitable for blood cell contour segmentation, and the improved method can be divided into two parts. In the first part, we use a weakly labeled dataset with the labeling type of bounding boxes for pretraining and then use the labels of the segmentation type for transfer learning of the cell segmentation model. In the second part, we add a smoothing constraint loss to the loss function of the mask to smoothen the segmented cell contours. We add the SE attention mechanism in the backbone network (ResNet18) to further improve the segmentation accuracy. Our method is mainly used for the segmentation of blood cell (erythrocyte and leukocyte) contours. Our method improves average precision (AP) by 8.4% and AP50 by 0.6% compared with PolarMask. The most significant improvement is in AP75, which improves by 8.8%. Our method models blood cell contours based on PolarMask and uses a weakly labeled training model to obtain pretrained weights that can segment red and white blood cells. Our method effectively improves the accuracy of the model in segmenting blood cells, and the segmented blood cell contours are smoother.