Research On Urinary Sediment Detection Based On Attention Mechanism And YOLOv5

Urinalysis is an essential component of routine medical examinations,enabling effective screening of major diseases such as urothelial carcinoma,renal tumors,and uremia through imaging and object detection of urine sediment.Currently,traditional machine learning-based urine sediment detection methods suffer from low accuracy,limited object recognition,and insensitivity to small objects.With the growing maturity of next-generation artificial intelligence technologies,deep learning-based object detection has achieved significant success in various fields,including agriculture and industry.In this context,this dissertation combines the latest deep learning theories and models to investigate urine sediment detection methods based on attention mechanisms and convolutional neural networks.The main contributions are as follows:1.Construction of a large-scale standardized dataset for urine sediment microscopic image.Currently,there is a lack of publicly available standardized datasets for urine sediment detection in the medical field,making it challenging for most deep learning models to be effectively applied.Therefore,this dissertation constructs a urine sediment dataset,named Urised2021,which contains over 30,000 images belonging to 12 different classes of urine sediments.The dataset’s diversity,object density,and relative object sizes are thoroughly analyzed.A comparison with existing publicly available datasets demonstrates the advantages of Urised2021 in terms of the variety of urine sediments,dataset size,and abundance of small objects.2.Design and implementation of a novel deep neural network for urine sediment detection.To address the limitations of the receptive field in object feature extraction using convolutional methods,several commonly used attention mechanisms are systematically analyzed in terms of their advantages and limitations in achieving global receptive fields and improving the performance of convolutional neural networks.Based on this analysis,a new attention module,named CSCA(channel spatial coordinate attention),is designed to enhance the YOLOv5 network for urine sediment detection.This algorithm achieves comprehensive feature extraction and accurate detection of urine sediments.Additionally,considering the issue of inconsistent feature scales in commonly used feature fusion networks,a detection head module based on adaptively all feature feature share(AAFS)is proposed.It employs self-learning to share spatial and channel-differentiated feature information from all feature layers except the current layer,ensuring scale invariance of features and effectively suppressing background information,thereby improving detection accuracy.