On Detection Algorithms For Formed Elements In Urine Sediment Micrograph

Urine sediment detection is one of the routine testing items in hospitals.It refers to the use of microscopes to detect the sediments of urine after centrifugation,and it is of great significance to provide quantitative indicators for the determination of related diseases.Traditional manual microscopy has a large workload,complicated operation and low efficiency,and it is easy to miss and misdetect.In recent years,the automatic detection of urine sediment based on image segmentation technology has become a research hotspot.However,due to the uncertainty of the acquisition process of microscopic images,the quality of microscopic images of urine sediment is often low,accompanied by more noise and adhesion.It is difficult to obtain accurate segmentation using image segmentation so as to affect the subsequent feature extraction and classification recognition.Based on this,machine learning methods are used to study the detection algorithm of urine sediments.Based on the dataset,the detection of red blood cells and white blood cells are mainly studied.The work is presented as follows:1.Relying on the project,pre-treating the microscopic images of urine sediment provided by the company.The dataset is made by referring to the book“Atlas of Urine Formed Element”and related professionals,including image annotation,single sample cutting and data augmentation.2.An ACF+ detector based on Aggregated Channel Features(ACF)is proposed to transform the traditional image segmentation plus feature extraction plus classification recognition task into a detection task.Using the aggregated channel features and its variants,combined with the Adaboost classifier which is based on decision tree,different detectors are designed for different urine sediments.Experiments verify the effectiveness of the proposed algorithm.3.An algorithm based on Imbalance Local Fisher Discriminant Analysis(ILFDA)for urine sediment detection is proposed.The algorithm uses the designed Haar-like template to filter the aggregated channel features(ACF)to extract the middle layer features,and to perform random subspace projection on the single-channel intermediate layer feature to construct multiple local multi-popular structures.For each locality,considering the imbalance of the training samples and the popular structure of the sample distribution,the ILFDA is proposed to learning weight coefficients to construct new features.4.A two-stage urine sediment detection algorithm based on SVM and Trimed MobileNets is proposed.HOG+SVM is used as region proposal module,considering the application of deep learning algorithms in the terminal,the requirement of speed and the importance of memory control,trimmed MobileNets is used as the classification module.The experiment proves that the proposed method is the smallest on model size and the detection speed is the fastest while ensuring the accuracy.5.Exploiting the urine sediment detection algorithm based on Faster R-CNN and analyzing the impact of different basic networks,anchor settings and pooling methods on detection performance.The proposed Faster R-CNN model is more suitable for urine sediment detection and achieves the best performance.The proposed urine sediment detection algorithms have a high accuracy and high speed,which has certain significance for promoting the development of automatic urine sediment detection.