Fluorescence Pattern Recognition Of ANA Based On Multi Instance Learning

Autoimmune disease refers to a disease caused by the immune system’s immune tolerance to its own components being reduced or destroyed due to some reasons,causing autoantibodies or sensitized lymphocytes to damage their own organs and tissues.Antinuclear Antibodies(ANA),as the most common type of autoantibodies in patients with autoimmune diseases,have important clinical significance for the classification,identification,typing,prediction,prognosis,and prevention of related diseases.The traditional ANA detection is mainly manual operation,which requires manual interpretation of the ANA image under the fluorescence microscope and judgment of the fluorescence pattern it belongs to.However,this method is time-consuming and labor-intensive,and the interpretation results are affected by subjective experience.In order to solve the problem of automatic recognition of ANA fluorescence patterns,this paper introduces a multi-instance learning method on the basis of convolutional neural networks,and formulates corresponding improvement schemes for the problems existing in the multi-instance learning method.The main work of this paper is as follows:1.For the ANA fluorescence pattern recognition task,this article introduces the commonly used convolutional neural network and the latest neural network for multilabel classification tasks,and tests their effects on the ANA dataset.2.This paper innovatively introduces a multi-instance learning method for the ANA fluorescence pattern recognition task,and divides it into a multi-instance learning method based on bag training and a multi-instance learning method based on instance training.The experimental results prove the effectiveness of the multi-instance learning method on the ANA dataset.3.The multi-instance learning methods based on instance training have the problem of untrustworthy instance labels,which causes the model to be disturbed by noise during training.Therefore,from the perspective of prior knowledge and instance entropy,this paper estimates the credibility of the instance label,and converts the estimated value into the sampling probability of the instance during model training.The experimental results prove that the improvement schemes based on prior knowledge and instance entropy respectively alleviate the problem of untrustworthy instance labels to a certain extent,and improve the performance of the model.4.The improvement scheme based on instance entropy cannot be cold-started,and the improvement scheme based on prior knowledge is relatively rough,so this paper combines these two schemes.The experimental results prove that the fusion scheme overcomes the inherent shortcomings of the respective schemes,better alleviates the problem of untrustworthy instance labels,and effectively improves the model effect.