Few-Shot Cervical Cell Images Instance Segmentation Based on Adversarial Training

The limited size of the accurately annotated cervical cell image datasets brings great challenges to the research of deep learning segmentation algorithms. This paper proposes an efficient instance segmentation method for fewshot cervical cell images. Firstly, find the ROI (region of interest) centered the nucleus for each cell. Then, the cells at the center of the ROIs are segmented by center cell segmentation network (CCSN) which is trained by adversarial mechanism, this net can learn cell morphology probability distribution from limited number of samples. In addition, we designed a dual-scale discriminator and applied it to adversarial training to improve the integrity of segmentation by strengthening the judgment of the center cell segmentation results. This paper conducts ablation and comparative experiments on the CCEDD dataset, when only 10 cervical cell TCT images participated in the training, the Dice of the proposed method reaches 92%, and the FNRo is improved by 32% compared with U-Net. Experimental results show that compared with existing methods, our method has better indicators under training with limited number of samples and can achieve accurate segmentation of cervical cell instances.