Towards a safe and efficient clinical implementation of machine learning in radiation oncology by exploring model...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiation Oncology

Background:

• Machine learning (ML) adoption is increasing in medicine, particularly in radiation oncology, leveraging advancements in deep learning and data availability.
• Radiation oncology has a history of digital workflows, making it receptive to ML integration.
• ML models, unlike traditional ones, are complex and data-dependent, raising concerns about performance and transparency.

Purpose of the Study:

• To review the risks and solutions for applying ML in radiation oncology workflows.
• To define and illustrate core concepts of interpretability, explainability, and data-model dependency in ML.
• To discuss ML applications and vendor perspectives for clinical integration.

Main Methods:

• Review of current literature on ML risks and solutions in radiation oncology.
• Formal definition and exemplification of interpretability, explainability, and data-model dependency.
• Discussion of ML applications and vendor viewpoints.

Main Results:

• ML implementation in radiation oncology is hindered by data-model dependency and lack of interpretability.
• Understanding and mitigating risks related to data and model interactions is essential.
• Key ML applications and vendor considerations for clinical use are explored.

Conclusions:

• Developing robust ML tools for radiation oncology requires addressing interpretability and data-model dependency.
• Risk assessment and quality assurance frameworks are needed for safe clinical deployment.
• Further research and collaboration are necessary for successful ML integration into radiation oncology practice.