In a study accepted by the International Conference on Computer Vision and Pattern Recognition, a research team led by Prof. WANG Quan and Prof. HU Bingliang from the Xi'an Institute of Optics and Precision Mechanics of the Chinese Academy of Sciences proposed a new algorithm named Beta-Fast Fourier Transform (β-FFT), which can provide stronger technical support for medical image diagnosis. 

In recent years, co-training has received considerable attention in the field of semi-supervised learning for its ability to utilize unlabeled data effectively to enhance model generalization. However, one major challenge for co-training approaches is the homogenization issue caused by models converging towards similar decision boundaries. 

In practical applications, due to similarities in model architecture, training data, and optimization algorithms, it often leads to these models gradually converging to similar decision boundaries, weakening their intended complementarity and affecting the generalization ability of the final model on unseen data.

In this study, researchers proposed a new algorithm named β-FFT with innovations in both data processing and training structure.

In terms of data processing, a nonlinear interpolation method based on FFT was employed to generate diversified samples by exchanging low-frequency components between images processed differently, enhancing model generalization and maintaining the stability of co-training. 

In the training structure aspect, a differentiated training strategy was designed where one model undergoes additional training using labeled data within a co-training framework, and linear interpolation following the β distribution is applied to unlabeled data as a regularization term. 

This strategy efficiently utilized limited labeled data, significantly improved the performance of the model on unlabeled data, enhanced the overall segmentation accuracy of the system, and performed at a world-class level on multiple public medical image datasets.

“This work provides an effective solution to the homogenization problem in semi-supervised learning and showcases its significant potential in the field of medical image diagnosis,” said Prof. WANG.