Detection And Segmentation Of Glands In Histopathological Images Of Human Colon Cancer Based On Deep Learning

Automatic and accurate segmentation of glands in histological images plays an important role in cancer diagnosis.Although the neural network-based gland segmentation method is better than the traditional gland segmentation method,the feature extraction in the traditional method can improve the neural network segmentation effect.All the experiments in this paper were tested on the Warwick-QU dataset of gland segmentation challenges announced at the MICCAI 2015 conference and compared with the latest research results.Aiming at the problem of "adjacent" gland adhesion,this paper proposes the Faster R-CNN model to detect glands.But the Faster R-CNN model is less effective in detecting malignant glands,In response to this problem,this paper proposes a new method that uses color features to improve the accuracy of Faster R-CNN in detecting malignant glands.This paper uses a probabilistic color detection model to extract color features.This paper uses a probabilistic color detection model to extract color features.Firstly,we expands the data set,then pre-processes the expanded data set,including extracting color features and histogram enhancement,and finally uses the pre-processed image to train the Faster R-CNN model.Experiments show that the new method proposed in this paper detects the F1-Score of malignant glands by 8.4%higher than the Faster R-CNN model without pretreatment.In the gland detection of the entire data set,the average F1-Score of the method proposed in this paper is improved by 3.9% compared with the Faster R-CNN model without preprocessing.Compared with the previous methods studied on the same data set,the F1-Score of the method proposed in this paper ranks in the top ten.At the same time,this paper also compares the experimental results of 6 preprocessing methods and 4data set enhancement methods.Aiming at the problem of single task neural network's weak segmentation ofgland contour,this paper proposes a multi-result fusion neural network model.The results of model fusion include gland region segmentation results,gland contour detection results and gaster detection results of Faster R-CNN in the previous study.U-net neural network works well in object contour segmentation,so we uses U-net for gland contour segmentation.First,for the problem that the single-task U-net neural network has a poor detection effect on the gland area,this paper proposes a new method to detect the gland region by combining the gland's compositional features,namely the cavity feature and the nuclear wall feature,and forming a three-task alternating neural network with the gland region.The loss function of each task branch of the model is independent,and the branch parameters are adjusted in order.Then,a single task U-net is used to detect the gland contours,and a simple neural network is used to fuse the three results to form the final gland segmentation result.Experiments show that when detecting gland regions on the same data set,the multi-task alternating model has a 9.2% improvement over the F1-Score of the single-task U-net model.In the segmentation of gland contours,MFSN-1 method using multi-task alternating model to detect gland regions,U-net contour segmentation and Faster R-CNN target detection results and use U-net gland region and contour segmentation results with Faster Compared with F1-Score,the MFSN-2 method of R-CNN target detection fusion is improved by 0.1%,and Object-level Dice is improved by 1.9%.At the same time,this paper also designed three different multi-task neural network models to detect the gland area,and conducted experimental analysis and comparison.