A few-shot u-net learning framework for fast and accurate three-dimensional dose prediction in radiotherapy

Accurate and fast patient-specific dose prediction is crucial to improve the quality and efficiency of radiation treatment planning. Recently, deep learning-based approaches have achieved remarkable results in dose predictions across various cancer types. However, they typically require a large set of training examples to ensure robust generalization.

To the end, our proposed FS-UNet leverages the U-Net framework and enhances dose prediction performance from limited training data using the discriminative power of few-shot learning. To achieve this, our FS-UNet is decomposed into three seamless components: (1) a 3D U-Net is a convolutional neural network specifically designed for dose prediction; (2) a gradient-based method MAML with N-way K-shot settings is embedded in the dose regression problem, which is a meta-learning algorithm by updating model's parameters based on a limited set of training samples; and (3) a Prototypical Network is utilized to perform an l-nearest centroid regression model to establish a feature metric generalizable for FS-UNet in dose prediction.

The FS-UNet model was validated for prostate proton radiotherapy using data from 102 patients, demonstrating superior performance against four existing deep learning-based approaches with both large (81 prostate cases) and small (11 prostate cases) training datasets. Additionally, experiments on 29 liver cases and 30 pancreas cases showed that dose prediction performance improved with mixed-site dose prediction compared to single-site predictions, even in few-shot learning scenarios.

The FS-UNet demonstrated more accurate DVH and less MSE in dose prediction for prostate, liver, pancreas, and mixed treatment sites in radiotherapy.