Detecting outliers beyond tolerance limits derived from statistical process control in patient-specific quality...

Area of Science:

• Medical Physics
• Radiotherapy Planning
• Machine Learning in Healthcare

Background:

• Patient-specific quality assurance (QA) identifies dose discrepancies in radiotherapy plans.
• Out-of-tolerance plans detected during measurement can cause significant treatment delays, often necessitating replanning.
• Early identification of potential plan deviations during the treatment planning phase is crucial to mitigate these risks.

Purpose of the Study:

• To develop and validate an outlier detection model for early identification of out-of-tolerance radiotherapy plans.
• To reduce treatment delays and the need for replanning by detecting plan issues during the planning stage.
• To improve the efficiency and reliability of radiotherapy treatment planning.

Main Methods:

• Utilized patient-specific QA data from portal dosimetry for stereotactic body radiotherapy (2020-2021).
• Analyzed data for thorax and pelvis sites using gamma passing rates (2%/2mm, 2%/1mm, 1%/1mm criteria).
• Employed statistical process control to establish tolerance limits and trained robust covariance, isolation forest, and one-class SVM models on plan complexity metrics.

Main Results:

• Defined site- and criterion-specific tolerance and action limits for pelvis and thorax sites.
• The one-class support vector machine model demonstrated superior performance.
• The best model achieved high recall (1.0) and F1-scores of 0.72 (thorax, 2%/2mm) and 0.70 (pelvis, 2%/1mm).

Conclusions:

• The developed outlier detection model enables early identification of potentially out-of-tolerance plans.
• Refinement of plans during the planning stage is facilitated, avoiding late-stage discoveries during measurement.
• This approach enhances treatment planning efficiency and reduces the risk of treatment interruptions.