Medical image segmentation algorithm based on positive scaling invariant-self encoding CCA

Abstract

The quality of medical image segmentation results directly affect disease analysis and diagnosis. Although traditional medical image segmentation methods have achieved certain results, this type of method is not only computationally inefficient, but also difficult to achieve fully automatic segmentation. Deep learning has a good generalization ability, which provides a new technical approach to solve the above problems. However, the following problems exist in the application of deep learning in medical image segmentation: (1) Optimization of deep learning models. The parameters of the existing deep learning model have not been well optimized; it is easy to cause wrong medical image segmentation. (2) Overfitting problem. As the number of deep learning model layers increases, it makes the deep model fit the training data easily, but its generalization performance on the test set will be poor. In view of this, in order to solve the problem of deep learning model optimization, this paper first reveals that the weight space of deep learning is not equivalent to its parameter space, and explains in detail how to optimize the deep learning model in a positive scaling invariant space. The method of mapping the derivative of the base path back to the weight, and analyzed the stochastic optimization algorithm in F space. In addition, to solve the overfitting problem of deep learning models, this paper first gives a method of self-encoding Canonical Correlation Analysis (Self-encoding CCA) to optimize both self-coding loss and correlation loss in the fine-tuning stage. Based on the above ideas, this paper proposes a medical image segmentation algorithm based on positive scaling invariant-self encoding CCA. The experimental results show that the index obtained by the method proposed in this paper is not only improved by a large margin compared with traditional machine learning methods, but also improved to a certain extent compared with other deep learning methods.