Cervical Cell Image Segmentation Based on Improved the Mask R-CNN Model

In the fields of medicine and biology, cell detection and segmentation play a vital role, providing powerful tools for disease diagnosis, biological research, and drug development. There are complex problems such as adhesion and overlap between cytoplasm in cervical cell images, and existing cervical cell image segmentation methods still have deficiencies in accurately segmenting cytoplasm and nucleus. Through improvements to the Mask R-CNN model, the segmentation capability of cell images has been enhanced, enhancing its applicability for segmenting cervical cell images. The main improvements include: (1) introducing attention mechanism into residual network to improve the ability of feature extraction and representation learning; (2) improving feature pyramid network to increase the receptive field size and enhance contextual information in the feature map. On the publicly available ISBI2014 dataset, the experiment demonstrates that, compared to the initial Mask R-CNN model, the segmentation metrics of cervical cell images, mIoU, Dice coefficient, and AP, increased by 2.54%, 3.55%, and 2.83%, respectively.