Research On Low-resolution Fine-granularity Classification Method Of Urinary Red Blood Cells

At present,there are many patients with chronic kidney disease in China,and there is an urgent need to establish a chronic kidney disease prevention and treatment system to achieve early screening,treatment,and management.When the number of red blood cells in urine exceeds the limit,it is a phenomenon of hematuria.The morphological classification of red blood cells in urine is an important basis for doctors to diagnose the source of hematuria.The morphological analysis and examination of urine red blood cells are known as in vitro kidney biopsy,which is of great significance for medical department testing.At present,the morphological detection method for urinary red blood cells in medical institutions in China is artificial microscopy.This method has low efficiency,high consumption of medical resources,and high professional requirements for testing personnel.The accuracy of existing urine red blood cell morphology analyzers cannot meet clinical requirements.The main purpose of this article is to improve the classification accuracy and recall of low resolution urinary red blood cells.In response to issues such as low spatial resolution of urinary red blood cell images,insufficient number of datasets,difficulty in separating samples,and model redundancy,combined with the characteristics of urinary red blood cell image classification tasks,the research content of this article is as follows:(1)A super-resolution reconstruction method based on category perception loss is proposed to address the problem of difficulty in feature extraction caused by low spatial resolution and poor image quality in urine red blood cell images.Based on the inter class similarity of urine red blood cell images,using high-resolution images of the same class to assist in the reconstruction of low-resolution images is beneficial for improving the image quality of discriminative regions in the image,facilitating the subsequent fine-grained classification network to extract the characteristics of discriminative regions in the image,thereby improving the accuracy of the classification model.The effectiveness of this method has been demonstrated through image quality assessment,visualization of super-resolution networks,and evaluation of image classification results.(2)To further improve the classification accuracy of urinary red blood cells,the finegrained classification network is further optimized.Firstly,in response to the insufficient number of urine red blood cell datasets and combined with the characteristics of urine red blood cell images,an RBC-MIX data augmentation method was proposed to improve the classification performance of the model.Then,for the difficult samples in the urine red blood cell classification task,propose class penalty loss,additive penalty and multiplicative penalty are introduced into the cross-entropy loss,which effectively improves the compactness within the class and opens the distance between different classes.Finally,the knowledge distillation method was used to compress the model,ensuring existing accuracy and making it more lightweight.The accuracy and recall indicators on the urine red blood cell test set,as well as the visualization results of the classification network,indicate that this method can effectively improve the accuracy and recall of the urine red blood cell classification model,and obtain a more lightweight fine-grained classification model.