Semi-supervised learning for dose prediction in targeted radionuclide therapy: a synthetic data study.

Area of Science:

• Medical Physics
• Radiology
• Artificial Intelligence

Background:

• Accurate radiation dose estimation is vital for effective Targeted Radionuclide Therapy (TRT).
• Deep learning (DL) shows potential for dosimetry, but requires large labeled datasets, which are often unavailable in clinical settings.
• This limitation hinders the widespread adoption of DL-based dosimetry in TRT.

Purpose of the Study:

• To develop and evaluate a semi-supervised learning (SSL) framework for radiation dose prediction in TRT.
• To reduce the dependency on large-scale labeled datasets by leveraging readily available pretherapy PET data.
• To adapt and extend traditional classification-based SSL approaches for regression-based dose prediction.

Main Methods:

• Exploration of a semi-supervised learning (SSL) framework utilizing pretherapy PET data with a small subset of dose labels.
• Adaptation and extension of classification-based SSL methods for regression tasks in dose prediction.
• Development of a synthetic dataset simulating PET images and Monte Carlo dose calculations for validation.

Main Results:

• Evaluation of multiple regression-adapted SSL methods under varying proportions of labeled data.
• Achieved overall mean absolute percentage errors of 9%-11% for dose prediction across different organs.
• Demonstrated performance comparable to fully supervised methods.

Conclusions:

• Proposed SSL methods show promising results for organ-level dose prediction in TRT.
• SSL effectively addresses the challenge of limited labeled clinical data for DL-based dosimetry.
• This approach enhances the feasibility of personalized dosimetry in TRT.