Evaluating dose distribution in prostate IMRT patients using deep learning: the influence of loss function on model...

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Deep learning (DL) models are increasingly used for radiotherapy dose prediction.
• Optimizing these models requires careful selection of loss functions.
• Prostate cancer intensity-modulated radiotherapy (IMRT) planning presents unique dosimetric challenges.

Purpose of the Study:

• To evaluate the impact of different loss functions on DL model performance for prostate IMRT dose prediction.
• To compare standard loss functions (MAE, MSE) with those incorporating domain-specific knowledge (DVH loss, moment loss).
• To assess the clinical applicability of DL-based dose prediction in IMRT.

Main Methods:

• Retrospective analysis of 110 prostate cancer patient cases.
• Training DL models with Mean Absolute Error (MAE), Mean Squared Error (MSE), MAE + Dose-Volume Histogram (DVH) loss, and MAE + moment loss.
• Performance evaluation using planned target volume (PTV) and organ at risk (OAR) dosimetric metrics.
• Statistical comparison using one-way analysis of variance (ANOVA).

Main Results:

• The DL model trained with MAE plus moment loss demonstrated superior performance in reducing dose deviations for OARs and PTV.
• Mean Absolute Errors (MAE) ± standard deviation (SD) for MAE + moment loss, MAE + DVH loss, MSE, and MAE were (1.76 ± 0.5) Gy, (1.78 ± 0.5) Gy, (1.93 ± 0.6) Gy, and (2.02 ± 0.4) Gy, respectively.
• All predicted dose distributions achieved clinically acceptable accuracy compared to ground truth plans.

Conclusions:

• Loss function selection is critical for optimizing DL-based prostate IMRT dose prediction.
• Integrating domain-specific knowledge into loss functions, particularly moment loss, significantly enhances model performance.
• These findings support the practical application of DL models for precise and personalized radiation therapy planning.