Cancer Cell Detection And Recognition Based On Deep Learning

With the increase of work pressure and the change of the environment,the rate of people in sub-health is getting higher and higher,which also leads to the higher and higher incidence of cancer.Therefore,cancer screening has always been a key examination item of physical examination.The method of investigation has important research value.Currently,assisted cancer screening relies mainly on pathologists,which is time-consuming and labor-intensive.With the application of computer vision in the medical field,the use of deep learning technology to detect and identify cancer cells for cancer screening has high research value.This paper studies how to use deep learning algorithms to identify and detect bladder cancer cells.The main contribution of the thesis includes the following three aspects:(1)Compressed U-Net algorithm based on PReLU:According to the characteristics of the experimental data set,this thesis designs a compression model based on the U-Net encoder-decoder structure,which compresses the original four upsampling and down-sampling steps to two Secondly,to obtain more efficient and accurate segmentation and recognition performance;in response to the problem of the ReLU activation function in the U-Net model,a more advanced PReLU method is introduced,which effectively alleviates the problem that the gradient is zero and the neurons cannot be activated.Compared with U-Net,the precision of pattern recognition based on compressed U-Net algorithm proposed by this thesis has a great performance improvement.(2)Faster R-CNN bladder cancer cell target detection algorithm based on K-means clustering algorithm:Aiming at the problem that the Anchor box in the Faster R-CNN network is not suitable for this data set,this thesis introduces the K-means clustering algorithm,through clustering calculation the size of the Anchor box suitable for the cancer cell data set in this experiment;then use the data set to train and test the improved algorithm network.In the experiment,different feature extraction models are used.After comparing the results of different experiments,it is found that the improved method combined with the average recognition accuracy of the ResNet-101 network has been significantly improved.(3)Improved YOLOv3 based on label smoothing regularization and the improved YOLOv4 based on dense connection target detection algorithm for bladder cancer cells:In view of the small number of bladder cancer cell data sets in this experiment,in order to improve the detection ability of the model,this article introduces several novel Image fusion algorithms,such as Mixup,Cutmix,and Mosaic methods.Aiming at the problem of potential label errors in the data set,in order to improve the model’s label error tolerance rate,this paper proposes a method of class label regularization to process the normal cell labels of cancer cells.Through different experimental results,it is found that the effect of the improved YOLOv3 original algorithm has been significantly improved;the improved YOLOv4 uses a densely connected form to replace the previous five convolutional layers that use the ReLU activation function in the neck module to output predictions in the previous step.The purpose is to make better use of image features.The results of ablation experiments show that the performance of the improved YOLOv4 algorithm in this paper is better than the improved Faster R-CNN and the improved YOLOv3 algorithm.