Combined Global And Region Of Interest Image Quality Assessment For Renal Pathology Based On Fusion Models

The quality of renal pathological sections not only has a decisive impact on the reliability of clinical pathological diagnosis,but also the quality of its digitized images may also affect the accuracy of remote diagnosis and computer intelligent analysis algorithms.First,high-quality renal pathological slides can effectively enhance the confidence of pathologists in the diagnosis results,thereby ensuring the accuracy of disease diagnosis and the rationality of subsequent treatment plans.Secondly,to mobilize remote diagnosis technology in areas where high-level expert service resources are lacking,in order to ensure the reliability of disease diagnosis,high requirements are also placed on the quality of digital pathological images that can be transmitted over long distances.In addition,the intelligent auxiliary diagnosis systems for renal pathology developed to reduce the workload of doctors also need to detect the quality of images before they are input into computer algorithms for intelligent classification and analysis.They train high-performance intelligent models through high-quality available images and obtain high reliability analysis results.At present,in actual pathological diagnosis and digital pathological image applications,the task of evaluating the quality of kidney pathological sections and digital images is mainly completed by pathologists through microscope observation.They need to repeatedly adjust the magnification of the slice and move the field of view under the light microscope,and subjectively judge its quality by their eyes.This process is not only time-consuming and labor-intensive,but also easily affected by individual observations,resulting in strong subjectivity and irreproducibility.When evaluating the quality of renal pathological images based on computers,the pathological images cannot be directly read and processed by computers due to their extremely high resolution and complex image features,which limits the development and application of automated evaluation methods for renal pathological image quality.Aiming at the above problems,based on the fusion deep learning model,this paper proposes a no reference renal pathological image quality assessment method that combines global and local regions of interest.Preprocessing data according to the clinical slice quality evaluation process,this method establishes a sub-image set for each high-resolution renal pathological image,effectively retaining the global and local fine-grained features of the renal pathological image,making the quality evaluation results more in line with the actual diagnostic requirements.The fused deep learning feature extractor can effectively deal with the high complexity of image features and ensure good performance of the model.In order to reduce the difficulty of obtaining quality score labels,this paper collects qualitative quality evaluation labels that are more in line with human psychology,and predicts both qualitative and quantitative quality conditions through algorithms.The method in this paper includes three modules: fusion CNN classification module,quality score calculation module and comprehensive quality evaluation module.The fusion CNN classification module realizes the qualitative classification of sub-images;the quality score calculation module,on the basis of calculating the sub-image quality scores,combines the global and region of interest quality to obtain the comprehensive quality score of the overall image to achieve quantitative evaluation;according to the fuzzy set theory,the comprehensive quality evaluation module disperses the continuous comprehensive quality scores into different quality sets to achieve qualitative quality evaluation.In this paper,kidney tissue whole slide images from 1105 patients were collected for training and testing of the quality assessment model.Through the trained model,the quality of a pathological image can be qualitatively evaluated end-to-end,and a quality score can be given at the same time.On the test set,the consistency between the method and the qualitative evaluation of senior pathologists can reach 90.05%,reaching the level of junior pathologists.At the same time,in the binary classification task of judging the availability of images,the accuracy can reach 99.095%.This shows that the method in this paper can more reliably evaluate the availability of renal pathological images in pathological diagnosis,remote diagnosis and intelligent analysis of renal pathology,and provide guarantee for subsequent pathological diagnosis and intelligent analysis.In addition,the quality evaluation database of renal pathological images established by this work contains a large number of high-quality and high-resolution images of glomerulus,which can be used for the development of lesions identification algorithms.And the established image quality evaluation criteria can also provide guidance for future image standardization and restoration.