Neural network dose prediction for cervical brachytherapy: Overcoming data scarcity for applicator-specific models

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• 3D neural networks aid in automating brachytherapy (BT) treatment planning for cervical cancer.
• Generalizing deep learning models across diverse BT applicators is challenging due to data variability and scarcity.

Purpose of the Study:

• To compare three neural network training methods: a single combined model, fine-tuning the combined model, and individual applicator models.
• To determine the optimal approach for accurate dose prediction in cervical brachytherapy.

Main Methods:

• A 3D Cascade U-Net was trained on 859 treatment plans across four applicator types (T&O, T&ON, T&R, T&RN).
• Methods compared included a combined model, fine-tuning the combined model, and individual (IDV) applicator models.
• Performance was evaluated using mean error (ME), mean absolute error (MAE), gamma analysis, and Dice similarity coefficients (DSC).

Main Results:

• Combined and fine-tuned models outperformed IDV models.
• Fine-tuning offered modest improvements in approximately half of the evaluated metrics.
• Low MAE and ME, high gamma pass rates (83-91%), and high DSCs (0.88-0.92) indicated accurate dose prediction.

Conclusions:

• Training on combined brachytherapy data enhances dose prediction accuracy, overcoming data scarcity for individual applicator types.
• Accurate, applicator-specific dose predictions can facilitate automated, knowledge-based planning for cervical brachytherapy.
• The combined dataset allows neural networks to learn generalizable dose trends across different applicators.