Available online 4 October 2022
Highlights•Re-parameterization network is used to recognize low-resolution microscopic images.
•Re-parameterization network can effectively extract the features of urine sediment images.
•CSAM increases the extraction of fine-grained features.
•Unbiased classifier is trained under the improved C.E. loss function.
AbstractObjectivesBecause there are many categories, large morphological differences and few labels of urinary sediment components, and uneven data in urine sediment images recognition, the accuracy and recall rate of the existing urine sediment images recognition methods are not ideal. The main purpose of this paper is to improve the accuracy and recall of urine sediment image recognition by proposing a urine sediment image classification method based on semi-supervised learning.
MethodsThis paper proposes a method based on semi-supervised learning to classify urine sediment images. This algorithm designs a re-parameterization network (US-RepNet) for low-resolution urine sediment microscopic images to extract complex features of urine sediment images. The dual attention module is introduced on Us-RepNet to increase the extraction of fine-grained features from urine sediment images. And the cross-entropy loss (C.E. loss) function is optimized to train an unbiased classifier to improve long-tailed distribution image classification.
ResultsThe experimental results show that the accuracy of proposed method can reach 94% with only a small amount of labeled data for 16 types of urine sediment images under long-tail distribution.
ConclusionThe algorithm can recognize most types, and reduces the need for labeled information, while achieving excellent recognition and classification performance. Comprehensive analysis shows that it can be used as a practical reference for urine sediment analysis.
Graphical abstractDownload : Download high-res image (65KB)Download : Download full-size imageKeywordsUrine sediment classification
Contrastive learning
Structural re-parameterization
Attention mechanism
View full text© 2022 AGBM. Published by Elsevier Masson SAS. All rights reserved.
留言 (0)