Improved outcome models with denoising diffusion.

Area of Science:

• Medical Physics
• Artificial Intelligence in Medicine
• Radiotherapy Research

Background:

• Radiotherapy outcome modeling frequently encounters challenges with imbalanced datasets for key endpoints.
• Class imbalance can hinder the accuracy and reliability of predictive models in radiotherapy.
• Generative models offer a potential solution for augmenting sparse datasets.

Purpose of the Study:

• To implement Denoising Diffusion Probabilistic Models (DDPM) for generating synthetic data to improve radiotherapy outcome models.
• To address class imbalance in tumor local control prediction models using a novel conditional 3D DDPM.
• To compare the performance of DDPM-augmented models against conventional class imbalance techniques.

Main Methods:

• A dataset of 535 NSCLC patients treated with SBRT was utilized for training.
• A conditional 3D DDPM was developed to generate synthetic radiotherapy planning data.
• Performance was evaluated by supplementing the real training data with synthetic data and comparing against Oversampling, Undersampling, Augmentation, Class Weights, SMOTE, and ADASYN.

Main Results:

• Synthetic data generated by DDPM were visually validated and achieved a Fréchet inception distance (FID) below 50.
• Augmenting the training dataset with DDPM-generated data led to over a 10% improvement in model performance.
• Conventional techniques provided only approximately 4% improvement in model performance.

Conclusions:

• DDPM presents an innovative method for tackling class-imbalanced outcome modeling in radiotherapy.
• The generated synthetic data are realistic and enhance the performance and robustness of outcome prediction models.
• This approach holds significant promise for advancing personalized radiotherapy treatment planning.