A 3D transfer learning approach for identifying multiple simultaneous errors during radiotherapy.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Deep learning models, including convolutional neural networks (CNNs), show promise for identifying treatment errors using dose comparison images.
• Clinical scenarios often involve multiple simultaneous anatomical and/or mechanical errors, posing a challenge for current error identification methods.

Purpose of the Study:

• To evaluate the capability of CNN-based error identification in complex scenarios with multiple simultaneous treatment errors.
• To assess the performance of a 3D CNN for multilabel classification of treatment errors at two distinct levels.

Main Methods:

• Simulated clinically realistic combinations of treatment errors in 40 lung cancer patients' plans and CT images.
• Trained a 3D CNN using a transfer learning approach on 2580 predicted 3D dose distributions, comparing them to error-free distributions.
• Evaluated an ensemble model of three CNNs and utilized relative dose difference as an alternative input metric.

Main Results:

• The CNN achieved high F1-scores for Level 1 error classification (main error type) but lower performance for Level 2 (error subtype).
• Using relative dose difference improved Level 2 performance, while an ensemble model reduced overfitting.
• Independent test set F1-scores were 0.86 for Level 1 and 0.62 for Level 2.

Conclusions:

• Simultaneous multiple errors in 3D dose verification data can be identified using deep learning models.
• CNNs demonstrate potential for enhancing accuracy and reliability in radiotherapy error detection.