Research On Gland Cancer Assistant Segmentation Based On Deep Learning

The incidence of diseases caused by glandular diseases such as colorectal adenocarcinoma,pancreatic cancer and breast cancer is increasing year by year.Adenocarcinoma has become one of the malignant tumors that seriously damage human health.At present,the diagnosis of glandular lesions mainly relies on screening by pathology,ultrasound,molybdenum target and nuclear magnetic resonance.In case of suspected cases,pathological puncture examination should be carried out,and molecular typing in pathological images should be analyzed to determine whether it is cancerous and the degree of malignancy.Therefore,the analysis of gland image structure size,shape and some other morphological manifestations is an important basis for the diagnosis of adenocarcinoma.The traditional pathological image analysis mainly relies on the manual observation and analysis of pathologists,but this manual analysis has a huge workload and high difficulty.Moreover,due to the different professional qualities of pathologists and their personal subjective experience,it will lead to inevitable misjudgments and poor reproducibility of results.In recent years,thanks to the emergence and development of full-section scanning equipment,pathologists are very convenient to analyze and save pathological pictures on the computer,which lays the foundation for computer-aided pathological picture analysis.In recent years,some auxiliary diagnostic techniques combined with image processing have emerged,which have effectively alleviated the difficulty of pathological image analysis.Therefore,this paper designs three segmentation algorithms according to image characteristics and task requirements:(1)A U-Net based colon gland pathological image segmentation algorithm was proposed.The algorithm is based on U-Net model,according to the characteristics of glandular cell segmentation data set,the network coding part is improved.Firstly,due to the large proportion of glandular cells in the whole segmented image,3×3 convolution cascade is used to increase the receptive field of the network.At the same time,the output feature maps of each convolution layer are spliced at the end to enrich the multi-scale information of the coding network and optimize the segmentation accuracy of small targets and boundaries.Finally,the compression and excitation module is introduced in the residual connection,and the useless information is suppressed by modeling the relationship between the channels,so as to optimize the initial transmitted feature map and reduce the interference to the coding network.(2)A two-branch glandular pathological image segmentation algorithm based on Transformer and CNN is proposed.Convolutional neural network has strong induction bias ability,and can still play a good performance when the data set is small.In addition,convolutional neural network has a strong advantage in image texture recognition.In contrast,Transformer has strong remote dependent capture capability,making it more prone to capture the shape of the target,which is what convolutional neural networks lack.Therefore,the algorithm in this chapter adopts the form of double-branch network to make convolutional neural network and Transformer network decode and encode in parallel.Then,in the encoding process,the feature graphs of the two are fused through the cross fusion module.By fusion,Transformer can alleviate the performance deficit caused by lack of pre-training,and integrate and complement the coding information obtained from the two to optimize its segmentation performance.Finally,the segmentation results of the two are combined through the segmentation header.(3)A semi-supervised nuclear segmentation algorithm based on consistent learning is proposed.Because the number of training images in nuclear segmentation data set is very small,it is hoped to introduce similar data sets for expansion.However,in the actual task,because of the different fields and direction of the task,the labels of the images are often different.Therefore,this chapter proposes a semi-supervised segmentation algorithm based on consistency learning,which uses similar data sets to discard unmatched image labels and takes images as unlabeled data to train the network.Based on the student-teacher model,this chapter uses the model proposed in the previous chapter as the network skeleton to extract the features in the image,and then uses the consistent learning strategy in the semi-supervised algorithm to train the network.In addition,by using teacher model and student model to generate false tags,and improving on the basis of the cross entropy loss function,the network is expected to impose certain penalties on the regions with low confidence in the prediction results.Through the improved loss function,the network is expected to give higher confidence prediction results for the unlabeled data.