BPPV nystagmus signal labeling method

Abstract

The invention relates to a BPPV nystagmus signal labeling method, and the method comprises the following steps: enabling an expert to label a part of selected high-quality video samples according to clinical diagnosis experience; according to labeled data of an expert, expanding a sample size to obtain a sufficient sample number by adopting data processing schemes such as data enhancement and positive sample expansion, and training a basic deep learning model as a Teacher Model; taking the unlabeled data as a test set, predicting the unlabeled data by using the Teaser Model, and giving a corresponding pseudo-label effect; selecting and abandoning the samples with negative false label display and low-confidence positive samples and independent positive samples; for continuous positive samples, annotating the samples with high confidence as positive samples, wherein the samples are used as new training samples. According to the method, the deep learning model is trained by using the limited annotation data set, and the new patient sample is automatically annotated, so that the annotation cost is reduced.

Description

[Technical field] [0001] The present invention relates to the model modeling of medical field, specifically a batch method of a BPPV eye shock signal. [Background technique] [0002] When modeling real video data, a certain amount of annotation data is required for supervisory learning. However, in the medical field, labeling samples often require professional doctors to participate, and the accuracy of the label is extremely high; at the same time, because of privacy, law and other related factors, the patient's real data often cannot be obtained from the outside, which leads to the hospital The tag's training sample has a very high cost. [0003] Depth learning model training requires a lot of annotation data, and this cost is very high in the medical field. For example, in the depth learning model proposed by the paper "Developing a Diagnostic Decision Support System for BenignParoxysmal Positional Vertigo Using A Deep-Learning Model", you need to label video data with tag vid...

AI Technical Summary

Problems solved by technology

When modeling for a tertiary hospital, there are less than 200 labeled video samples available, which has become the biggest obstacle to deep learning model training

And for non-professional data organizations, the cost of obtaining labeled samples is very high

[0004] In addition, after a detailed study of the etiology of BPPV symptoms and video samples, the following conclusions were drawn: (1) the number of positive samples (with nystagmus) was much smaller than that of negative samples; Phase motion, the signal on the time series of pupil trajectory is very clear; (3) The labeling cost of positive samples is higher than that of negative samples

Images