Semantic Segmentation Of Medical Cervical Cell Images Based On Multi-scale And Multi-input Convolution Neural Network

Cervical cancer is a kind of disease which threatens women's health with a high morbidity.The analysis of cervical cells is of great significance in helping the diagnosis of cervical cancer.The accuracy of microscopic cell segmentation can directly affect the discriminative diagnosis of diseases.Especially in the microscopic pathological images of cervical cells,the shape and size of the nucleus as well as the proportional parameters between the cytoplasms are of great significance for the diagnosis of the disease.In order to improve the accuracy of cervical nucleus segmentation,the following work has been done in this paper:(1)This paper made and calibrated a CCTCT data sets of cervical TCT cell smear based on the THP(Thinprep cytologic test)tableting technique of Changsha Second People's Hospital under the guidance of the pathologist in this hospital.By using the data expansion technology for cervical TCT microscopic cell images,which combines the image rotation reflection transformation,the scaling transformation,the translation transformation and the random noise transformation method,it solves a series of problems that possibly caused by fewer training samples(2)On the basis of the classical medical image segmentation network model U-net,this paper proposes a multi-scale and multi-input semantic segmentation network model CellSeg-net based on the image of medical cervical cells of convolutional neural network by improving the decoder and encoder structure of the segmentation algorithm.In the process of the decoder,due to the key role of the cell image's detailed information plays in the segmentation effect of the image,the author puts forward the method of multi-scale extraction of the microscopic cell image characteristics and combines the multi-input training method,making the loss of image detail information the lowest in this process.And in the process of the encoder,the final segmentation effect is optimized by fusing the feature maps of similar sizes.(3)Through the detailed simulation experiments,the contribution of each improvement to the overall segmentation effect is elaborated.A method of using the convolution operation to take the place of the pooling operation is also proposed by analyzing the impact of the pooling operation on the overall segmentation effect.Meanwhile,by adding the BN layer,the present study improves the gradient flowing through the network and enhances the training speed greatly.And the improved CellSeg-net network model is also put forward.In this paper,satisfactory findings can be obtained through the results of simulation experiments(MIoU value is 92.6%),which proves the feasibility of the algorithm.