Explainable one-class feature extraction by adaptive resonance for anomaly detection in quality assurance

Area of Science:

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

Background:

• Radiotherapy (RT) plan quality assessment (QA) is critical for cancer treatment safety.
• Traditional QA involves iterative expert review, leading to class imbalance issues for automated methods.
• Complexity of RT plans and data imbalance hinder traditional binary classification for automated QA.

Purpose of the Study:

• To develop a novel framework for automated radiotherapy plan quality assessment.
• To address the challenges of class imbalance and complexity in RT plan QA.
• To improve the efficacy of machine learning in classifying acceptable versus unacceptable RT plans.

Main Methods:

• Introduction of a novel one-class classification framework.
• Utilizing an adaptive neural network architecture for anomaly detection.
• Evaluating performance against traditional binary and standard one-class classification methods.

Main Results:

• The proposed one-class classification framework outperforms traditional methods in imbalanced RT plan QA.
• The method enhances anomaly detection capabilities without sacrificing interpretability.
• Demonstrated effectiveness in complex and imbalanced datasets inherent to RT plan QA.

Conclusions:

• The novel framework offers a more effective approach to automated RT plan QA.
• Enhanced interpretability facilitates healthcare professionals' understanding and trust in automated decisions.
• Streamlines the QA process, improving patient care efficiency and safety in radiotherapy.