Robust deep learning-based patient-specific quality assurance prediction models for novel dual-layer MLC linac

Area of Science:

• Medical Physics
• Artificial Intelligence in Healthcare
• Radiation Oncology

Background:

• Patient-specific quality assurance (PSQA) is crucial for safe and effective radiation therapy.
• Fixed field intensity-modulated radiation therapy (FF-IMRT) on advanced linear accelerators (linacs) presents unique challenges for PSQA.
• Accurate prediction of PSQA outcomes is essential for timely identification of potentially failing treatment plans.

Purpose of the Study:

• To evaluate the feasibility of deep learning models for robust PSQA outcome prediction in FF-IMRT on Halcyon linacs with dual-layer MLC.
• To enhance PSQA prediction accuracy and robustness using Shuffle Attention (SA) and deep imbalance regression.
• To ensure reliable predictions despite extreme imbalance in gamma passing rate (GPR) data.

Main Methods:

• Developed an ALDS-ResNet model integrating ResNet with SA and Label Distribution Smoothing (LDS).
• Trained and validated the model using data from 214 FF-IMRT plans and 1394 beam orientations.
• Utilized fluence maps as input for the deep learning models.

Main Results:

• ALDS-ResNet demonstrated significantly lower Mean Absolute Error (MAE) compared to standard ResNet across various gamma criteria (e.g., 1%/1 mm, 2%/2 mm, 3%/3 mm).
• The model showed improved performance on complex plans with low GPR values (GPR < 85).
• ALDS-ResNet achieved higher Pearson correlation coefficients (CC) at 2%/2 mm and 3%/3 mm gamma criteria compared to ResNet.

Conclusions:

• Deep learning models, particularly ALDS-ResNet, show strong potential for predicting GPR values in dual-layer MLC linacs.
• Integration of attention mechanisms (SA) and LDS significantly enhances PSQA prediction accuracy and robustness.
• The ALDS-ResNet model can serve as a valuable auxiliary tool for physicists in identifying PSQA failure plans, improving clinical workflow.