Multimodal radiotherapy dose prediction using a multi-task deep learning model.

Area of Science:

• Medical Physics
• Radiotherapy
• Deep Learning

Background:

• Accelerated partial breast irradiation (APBI) is preferred over whole breast irradiation for its targeted dose delivery and shorter treatment duration.
• Various APBI modalities exist, each with unique dose distribution characteristics, necessitating patient-specific optimal modality selection.
• Manual treatment planning for each modality is time-consuming and clinically impractical.

Purpose of the Study:

• To develop an efficient, personalized approach for selecting the optimal radiation therapy (RT) modality for APBI using deep learning.
• To train a multi-task (MT) network capable of concurrently predicting dose distributions for various RT modalities.
• To provide quantitative, patient-specific dose insights for informed modality comparison prior to treatment planning.

Main Methods:

• A dataset of 28 APBI patients and 92 treatment plans was used, partitioned into training, validation, and test sets.
• Single-task (ST) models were trained for each modality, alongside a single MT model predicting doses for all modalities.
• Model performance was evaluated using Mean Absolute Percent Error (MAPE) on the test dataset, with statistical analysis via Wilcoxon signed-rank test.

Main Results:

• The MT model required 2384 minutes for training, while five ST models totaled 1925 minutes.
• MT model predictions averaged 1.82 seconds per patient, compared to 0.93 seconds per modality for ST models.
• The MT model achieved a significantly lower MAPE (1.1033 ± 0.3627%) than the collective ST models (1.2386 ± 0.3872%).

Conclusions:

• A multi-task learning framework can predict RT dose distributions across modalities effectively without significant compromise.
• The MT architecture offers flexibility, scalability, and streamlined management, making it suitable for clinical deployment.
• This approach enhances patient decision-making, provides physicians with quantitative insights, and optimizes clinic resource allocation.