CTGAN in Augmentation of Radiomics Features Classification from Narrow Band Imaging for Laryngeal Cancer

Artificial intelligence (AI) holds immense promise in revolutionizing biomedical research, particularly in the field of early medical assistance analysis. This paper explores the application of AI in the context of laryngeal cancer, a disease where early screening, accurate diagnosis, effective management, and favorable prognosis are crucial for patient outcomes. The implementation of AI in biomedical field practice always faces challenges due to data scarcity. The limited availability of data leads to less satisfactory testing results. Even if the methods like geometric transformation and photometric transformation have been applied, it does not still enlarge the diversity in nature of data. Here, we investigated CTGAN on tabular radiomics features in a public laryngeal cancer dataset to check how various amount of data augmentation affects classifier's performance. The results were assessed by the synthetic data reports which captures the similarity with the columns shapes score (median value 71.23 %) and the trend and correction across columns with a column pair score median value 90.30%. The synthetic data respect the original data structure(100%) and overall synthetic data validity is above 81 %. It enhances the diversity and increase the amount of training data for laryngeal cancer detection. After assessing the synthetic report, a comparison of performances across different classifiers was followed. Result shows an increases in accuracy from 5 % to 10%. This proves the positive performance of the classifying improvement on an independent testing dataset (real data) and provides clues how much data should be synthesized. Our paper provides a positive and meaningful reference on tabular radiomics data augmentation for medical intelligent diagnosis design in the future.