Research On Semantic Information Based Zero-Shot Medical Image Classification Method

Medical image uses physical methods such as light and electricity to obtain images of the inside of the human body or a part of the human body in a non-invasive manner,and it is one of the most important bases for clinical medical diagnosis.Medical image data accounts for more than 90% of medical data,and it clearly occupies an important position in the medical diagnosis system.However,there are problems with data interoperability,scarcity,privacy,and other issues in the modern medical system.In this case,using Computer-Aided Diagnostic technology for automated classification is of great significance.Due to the ability of zero-shot learning to recognize new categories that have not been seen before,zero-shot medical image classification has tremendous potential in real-world scenarios.This paper proposes different methods to address the needs of zero-shot chest X-ray classification and zero-shot fundus color map classification from the perspective of the rich semantic information that medical images possess.The main research includes the following three points:Zero-shot chest X-ray classification based on visual attribute semantics.This paper first focuses on zero-shot chest X-ray classification and addresses two common issues:semantic gap between visual feature space and semantic space,and the inconsistency in image quality.To solve these problems,this paper proposes a variational autoencoder network model based on visual attributes,which is trained on a large public dataset and tested on multiple large public test datasets.Experimental results demonstrate the robustness and effectiveness of the proposed method.Zero-shot fundus color map classification based on class hierarchy semantics.This paper then shifts its focus to zero-shot fundus color map classification.To address the distortion problem caused by embedding the category hierarchy semantics in Euclidean space,this paper proposes a hyperbolic neural network model based on the category hierarchy structure,utilizing hyperbolic space to embed the category hierarchy structure.This paper also improves the zero-shot classification-related definitions to apply them to fundus color map classification.The experimental results demonstrate that the hyperbolic embedding method can obtain a more robust zero-shot fundus color map classification model.Finally,a self-training framework based on input noise is proposed to further reduce the influence of label noise in large-scale medical image data sets.The improved input noise-based self-training framework can effectively improve the noise in the dataset and self-training itself,thus improving performance.The experimental results demonstrate the effectiveness and generality of self-training and the input noise-based self-training framework on distance metric classification models.