Uncertainty Assessment for Deep Learning Radiotherapy Applications.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Deep learning (DL) has rapidly advanced in radiotherapy over the last 5 years.
• Radiotherapy data's structured nature facilitates DL applications, offering automation potential for laborious tasks.
• Fundamental weaknesses in DL, particularly regarding data origin assumptions and residual uncertainties, necessitate careful consideration for clinical use.

Purpose of the Study:

• To explain concepts of uncertainty quantification in DL for the radiotherapy community.
• To detail epistemic and aleatoric uncertainties and their modeling techniques in DL.
• To demonstrate the application of uncertainty assessment in radiotherapy DL workflows.

Main Methods:

• Review and explanation of general concepts of uncertainty in DL, drawing from computer vision literature.
• Detailed description of epistemic and aleatoric uncertainty modeling techniques.
• Demonstration using three radiotherapy DL applications: dose prediction, synthetic CT generation, and contouring.

Main Results:

• DL models must indicate when predictions exceed uncertainty thresholds for safe clinical deployment.
• Uncertainty quantification can enhance confidence in automated DL-driven radiotherapy workflows.
• While promising, current methods for distinguishing in-distribution from out-of-distribution samples are immature, requiring QA and human oversight.

Conclusions:

• Deep learning offers significant value in radiotherapy, but its clinical integration demands rigorous uncertainty assessment.
• Understanding and modeling epistemic and aleatoric uncertainties are key to maximizing DL's potential in radiotherapy.
• Continued research and development are needed for robust automated quality assurance in DL-based radiotherapy systems.