Overcoming data scarcity in radiomics/radiogenomics using synthetic radiomic features

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Radiomics and radiogenomics models are crucial for medical AI but often suffer from data scarcity.
• Limited availability of high-quality, annotated medical data hinders the development and validation of robust predictive models.
• Synthetic data generation offers a potential solution to augment real-world datasets.

Purpose of the Study:

• To evaluate the efficacy of synthetic radiomic data generation in addressing data scarcity for radiomics and radiogenomics models.
• To assess the impact of synthetic data augmentation on the performance of predictive models trained with limited sample sizes.
• To determine the quality and reproducibility of synthetic radiomic data generated using tabular models.

Main Methods:

• A retrospective cohort of 386 colorectal cancer patients with CT images and TP53 mutational status was used.
• Real-world radiomic features were extracted and used to train five different tabular synthetic data generation models.
• Subsets of training data (200, 400, 1000 lesions) were used to evaluate model performance with and without synthetic data augmentation.

Main Results:

• Models trained on limited real-world data showed poor predictive performance (AUC 0.52-0.56).
• Synthetic data generation models produced reproducible synthetic radiomic data highly similar to real-world data.
• Integrating synthetic data enhanced predictive model performance by 9.6% to 16.7% for small training sets.

Conclusions:

• Synthetic radiomic data, when combined with real data, can significantly enhance predictive model performance.
• Tabular synthetic data generation shows promise in overcoming data scarcity limitations in medical AI.
• High-quality synthetic data generation requires real-world signal data, not randomized or noisy data.