Hybrid Probabilistic Inference Based on Medical Image Attributes

Hekun Tian

Keywords: Causal reasoning; neural networks; mixed probability inference; medical image diagnosis

Abstract:

Deep neural networks have widespread applications in medical image diagnosis, and have achieved significant results in
tumordetectionandnoduleclassification. However, theperformanceofthemodeliscloselyrelatedtothequalityofthedataset. Without
a large and high-quality training data set, the neural network model is difficult to perform well through training. The most significant
probleminthemedicalfieldisthedifficultyof obtaining high-quality medical data sets, which makes it difficult for pure neural networks
to overcome the problems of poor generalization, lack of explainability, and poor verification performance. In view of these problems,
this paper proposes a hybrid probabilistic inference method, combining causal reasoning with neural networks to build a powerful
inference system. Specifically, the features of the pathological images need to be extracted first, represented as pathological attributes,
and then a graph neural network is used to establish relationships between attributes and diseases. A Bayesian network is used for
causal inference modeling, and the results of the two are tightly coupled to obtain a diagnosis method with strong generalization ability,
good verification effect, and explainability.This method can be used for CT image lung nodule classification, achieving an accuracy
of 95.36 precent; it can also be used to diagnose tuberculosis using chest x-ray images, achieving an accuracy of 96.64 percent. The
experimental results show that the hybrid probabilistic inference algorithm has a huge improvement in the performance of pure neural
network models, and provides a new idea for further research on causal reasoning and neural networks.

Submission Number: 13