What is the optimal input information for deep learning-based pre-treatment error identification in radiotherapy?

Area of Science:

• Medical Physics
• Radiotherapy Quality Assurance
• Machine Learning in Healthcare

Background:

• Pre-treatment quality assurance (QA) in radiotherapy is crucial for patient safety.
• Deep learning (DL) shows promise for enhancing the sensitivity of error detection in radiotherapy QA.
• Systematic evaluation of DL model performance requires understanding the impact of various input parameters.

Purpose of the Study:

• To systematically evaluate the influence of different dose comparison and image preprocessing techniques on deep learning (DL) model performance for identifying errors in pre-treatment radiotherapy QA.
• To compare the effectiveness of various dose comparison metrics and image preprocessing strategies for error detection.
• To determine the optimal parameters for maximizing DL model accuracy in radiotherapy QA.

Main Methods:

• Simulated mechanical errors (MLC leaf positions, monitor unit scaling, collimator rotation) in 53 VMAT and 69 SBRT lung cancer patient plans.
• Compared portal dose images using standard (gamma analysis), simple (absolute/relative dose difference, ratio), and alternative (distance-to-agreement, structural similarity index, gradient) dose comparison methods.
• Evaluated different normalization methods (min/max, mean/stdev) and image resolutions (32x32, 64x64, 128x128) for preprocessing.

Main Results:

• Simple dose comparison methods yielded the highest average accuracy (Level 1: 97.7%, Level 2: 78.1%), outperforming alternative methods (Level 1: 91.6%, Level 2: 71.2%).
• Mean/stdev normalization notably enhanced Level 2 classification accuracy.
• Increased image resolution generally improved error identification, though lower resolutions were adequate for SBRT.

Conclusions:

• The selection of dose comparison method is the most critical factor influencing deep learning-based error identification in pre-treatment radiotherapy QA.
• Optimizing DL model performance can be achieved by employing simple dose comparison methods, mean/stdev normalization, and high image resolution.
• These findings provide a pathway for improving the reliability and efficiency of automated radiotherapy QA processes.