Research On Digital Pathology Image Quality Detection Method Issue Based On Computer Aid

Digital pathology image with high quality is the prerequisite for the accurate application of digital pathology.In order to propose a more applicable method for detecting quality problems of digital pathology image,this paper divides the problem into slide-level and patch-level.Slide-level quality problems include incorrect focal length setting,cover slip misplacement and the derivation of virtual position and actual position.These problems will cause the whole-slide image unusable.Patch-level quality problems refer to unusable picture areas with poor visual effects in whole-slide.Aiming at the three slide-level quality problems,this paper proposes corresponding detection methods.First of all,the detection of the wrong focus setting depends on the information richness of the image.Since the Laplacian gradient function can well characterize the information richness of the image,this paper proposes a method based on the Laplacian gradient function to detect the problem.Secondly,cover slip misplacement is morphologically expressed as a linear object.Since the Frangi2 D filter feature has a significant effect of dividing linear objects,this paper proposes a method based on the Frangi2 D filter feature to detect this problem.Thirdly,the derivation of virtual position and actual position will result in the loss of clinically significant image areas.The edge image block contains a large number of non-background area blocks,which can indicate that the image has this problem.For this reason,this paper proposes the problem of detecting this problem through the proportion of the non-background areas of all edge areas.Aiming at the problem of patch-level quality,this paper proposes a classification method based on class incremental learning to solve the demand of dynamically increasing the number of classes that need to be detected.The class incremental learning method takes the pre-trained pathological image classification model for identifying human tissues as the initial state.Then,the method calculates the mean value of the feature vector of each category during the training process.In the category increment step,firstly we sample in the old category and use the convolutional neural network to extract the features,then we save the current average of the feature vectors of all categories;next we add a feature aggregation structure to store the feature vector information of the previous state Add to the final feature calculation result,and finally use the cosine normalization classification method to classify the obtained features.In the experimental part,this paper verifies the feasibility of the proposed method.Aiming at the slide-level quality problem,this paper designs an experiment to adjust the parameters of the independent variables in the method to obtain the best detection method;and proves the effectiveness of each method in the validation dataset.Aiming at the problem of patch-level quality,this paper designs two sets of comparative experiments and ablation experiments.This paper compares the proposed class incremental learning method with existing detection methods and general class incremental learning methods.This verifies that the patch-level quality detection method based on class incremental learning proposed in this paper is superior to existing methods in performance.Finally,this article proves the effectiveness of the method design through ablation experiments.