On December 13, a research group led by Graduate Student Hiroki Oka, Lecturer Daisuke Kawahara, and Professor Yuji Murakami of the Department of Radiation Oncology, Graduate School of Biomedical and Health Sciences at Hiroshima University announced that they developed a new technique to correct imbalances in medical images in prognosis prediction by machine learning using radiomics analysis. They confirmed that using a Gaussian noise-based imbalance correction method (GNUS) could improve prediction accuracy by reducing the bias toward the majority in artificial intelligence (AI) prediction, which is caused by the use of biased data, for example, those dominated by specific cases. The findings are expected to improve the accuracy of predicting postoperative recurrence in cancer patients. The results were published in the international journal Computers in Biology and Medicine on July 26.
Recently, AI analysis of medical data has been used to predict and detect diseases early and optimize treatment methods. Radiomics analysis, in which characteristics unnoticeable by human eyes are extracted from medical images, and AI learns them to make advanced predictions, has been attracting particular attention. Meanwhile, medical image analysis faces the challenge of being unable to make accurate predictions for minor cases because the proportion of cases to be analyzed is biased. The AI thus prioritizes predictions based on data from the majority of cases. Oversampling methods have been developed to solve this weakness, but correcting multidimensional data containing many features (e.g., age, blood test values) has remained difficult.
To overcome this problem, the research group developed a data imbalance correction method by hypothetically increasing the number of data in minority groups using Gaussian noise in each dimension of the ultra-multidimensional data and examined whether the prediction accuracy could be improved. Gaussian noise is stochastic noise whose values follow a Gaussian distribution (normal distribution) and is used for noise modeling. The effectiveness of GNUS in predicting recurrence of head and neck cancer using computed tomography (CT) and positron emission tomography (PET) images and in predicting recurrence in patients with squamous cell carcinoma of the head and neck was tested.
The results showed that the AI increased overall prediction accuracy while reducing the tendency to be biased toward the majority. The system developed had an accuracy level sufficient for clinical application. Moving forward, the research group aims to validate the system using data from other facilities and develop applications.
Kawahara said, "In AI prediction of prognosis using radiomics analysis, such as the analysis in this study, about 10 image features with reduced dimensionality are input to the AI instead of all image features. As the problem of the conventional data imbalance correction is the use of data irrelevant to prediction, we focused on solving this problem. In medical AI research that deals with images, it is important to perform imbalance correction, like our method, that incorporates specified types of data instead of using unnecessary data, and we expect that our method will be utilized in the future."
Journal Information
Publication: Computers in Biology and Medicine
Title: Radiomics-based prediction of recurrence for head and neck cancer patients using data imbalanced correction
DOI: 10.1016/j.compbiomed.2024.108879
This article has been translated by JST with permission from The Science News Ltd. (https://sci-news.co.jp/). Unauthorized reproduction of the article and photographs is prohibited.