The structural similarity index for IMRT quality assurance: radiomics-based error classification

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Healthcare

Background:

• Conventional gamma analysis in intensity modulated radiotherapy (IMRT) quality assurance (QA) has limitations in accurately identifying delivery errors.
• Radiomics and machine learning (ML) show promise for enhancing error detection in IMRT QA.
• Structural SIMilarity (SSIM) sub-index maps can reveal specific types of dose distribution errors.

Purpose of the Study:

• To apply radiomics analysis on SSIM sub-index maps for feature extraction.
• To develop and validate ML models for classifying machine-related delivery errors in patient-specific dynamic IMRT QA.
• To compare the performance of the ML-based method against conventional gamma analysis.

Main Methods:

• Simulated four types of machine-related errors (MU variations, MLC shifts, random MLC mispositioning) in 21 IMRT plans.
• Acquired 1620 portal dose images and calculated difference maps, including SSIM sub-index maps.
• Extracted radiomic features, employed SVM recursive feature elimination for feature selection, and tested classifiers (LDC, SVM, RF) using nested cross-validation.

Main Results:

• The Linear-SVM model achieved the highest overall classification accuracy of 0.86.
• Average classification accuracies for shift, opening, and random errors were approximately 0.9.
• The ML model demonstrated superior performance compared to gamma analysis (2 mm/2% criterion), with higher sensitivity to errors.

Conclusions:

• An ML-based method utilizing radiomics on SSIM sub-index maps effectively identifies machine-related errors in dynamic IMRT QA.
• The developed ML models achieved high classification accuracies, surpassing conventional gamma analysis.
• This radiomics and ML approach holds significant potential for improving error detection in patient-specific IMRT QA.