Identifying the optimal deep learning architecture and parameters for automatic beam aperture definition in 3D...

Area of Science:

• Medical Physics
• Artificial Intelligence in Healthcare
• Radiotherapy Technology

Background:

• Two-dimensional radiotherapy is a common cervical cancer treatment in low- and middle-income countries.
• Treatment planning for 2D radiotherapy is complex and time-consuming.
• Neural networks can automate planning, but hyperparameter impact on accuracy needs investigation.

Purpose of the Study:

• To evaluate the effect of convolutional neural network architectures and hyperparameters on 2D radiotherapy treatment field delineation.
• To identify optimal hyperparameters for accurate automated treatment planning.

Main Methods:

• Trained six deep learning architectures to delineate four-field box apertures on digitally reconstructed radiographs.
• Conducted a comprehensive hyperparameter search, varying learning rate, normalization, kernel size, network depth, feature maps, and activation functions.
• Trained and tested over 1700 unique models.

Main Results:

• Initial learning rate of 0.0001 was consistently significant for improved performance.
• DeepLabv3+ and D-LinkNet architectures demonstrated the least sensitivity to hyperparameter selection.
• Achieved high overlap (mean Dice similarity coefficient = 0.98) and surface distance agreement (mean surface distance < 2 mm).

Conclusions:

• DeepLabv3+ and D-LinkNet are the most robust architectures for hyperparameter selection in this task.
• Learning rate, nonlinear activation function, and kernel size are key hyperparameters for enhancing performance.