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

Abstract

Purpose: Two-dimensional radiotherapy is often used to treat cervical cancer in low- and middle-income countries, but treatment planning can be challenging and time-consuming. Neural networks offer the potential to greatly decrease planning time through automation, but the impact of the wide range of hyperparameters to be set during training on model accuracy has not been exhaustively investigated. In the current study, we evaluated the effect of several convolutional neural network architectures and hyperparameters on 2D radiotherapy treatment field delineation.

Methods: Six commonly used deep learning architectures were trained to delineate four-field box apertures on digitally reconstructed radiographs for cervical cancer radiotherapy. A comprehensive search of optimal hyperparameters for all models was conducted by varying the initial learning rate, image normalization methods, and (when appropriate) convolutional kernel size, the number of learnable parameters via network depth and the number of feature maps per convolution, and nonlinear activation functions. This yielded over 1700 unique models, which were all trained until performance converged and then tested on a separate dataset.

Results: Of all hyperparameters, the choice of initial learning rate was most consistently significant for improved performance on the test set, with all top-performing models using learning rates of 0.0001. The optimal image normalization was not consistent across architectures. High overlap (mean Dice similarity coefficient = 0.98) and surface distance agreement (mean surface distance < 2 mm) were achieved between the treatment field apertures for all architectures using the identified best hyperparameters. Overlap Dice similarity coefficient (DSC) and distance metrics (mean surface distance and Hausdorff distance) indicated that DeepLabv3+ and D-LinkNet architectures were least sensitive to initial hyperparameter selection.

Conclusion: DeepLabv3+ and D-LinkNet are most robust to initial hyperparameter selection. Learning rate, nonlinear activation function, and kernel size are also important hyperparameters for improving performance.

Keywords: automatic segmentation; cervical cancer; digitally reconstructed radiograph; radiotherapy.

FIGURE 1

Four‐field conversion process for the 310 digitally reconstructed radiographs (DRRs) used in our study. AP, anterior‐posterior; CT, computed tomography; PA, posterior‐anterior; RL, right‐lateral; LL, left‐lateral.

FIGURE 2

Dice similarity coefficient (DSC) values for each architecture (rows) by learning rate and image intensity normalization scheme (columns). Normalization schemes are abbreviated as follows: a, z‐score normalization with global values (μ = 25 and σ = 30); b, z‐score normalization with individual image values; c, histogram stretching (I_bottom  = 25, I_top  = 60); d, histogram stretching (I_bottom  = 0, I_top  = 45); e, histogram stretching (I_bottom  = 0, I_top  = 90); f, histogram stretching (I_bottom  = I_min , I_top  = I_max ); g, L₂ normalization. The middle column (Normalization) contains results for models trained with all learning rates, while the right column (Normalization*) includes only results for models trained with a learning rate value of 0.0001.

FIGURE 3

Effect of hyperparameters selected for U‐Net architectures on Dice similarity coefficient (DSC) values in the test set. (a) Comparison of traditional U‐Net and Res‐U‐Net, which employs residual connections. (b) Comparison of ReLU and PReLU activation functions. (c) Comparison of concatenation and element‐wise addition of features in the residual connections of the Res‐U‐Net.

FIGURE 4

Effects of field angle on Dice similarity coefficient (DSC) values in the test set. For all architectures, median DSC values were lower for right‐lateral (RL) fields than for anterior‐posterior (AP) or posterior‐anterior (PA) fields. For clarity, only results for models trained with a learning rate of 0.0001 are included; the general trend was the same for all models regardless of learning rate.

FIGURE 5

Examples of predictions with “good” and potentially problematic anatomic features in a top‐performing model. All predictions were made using the DeepLabv3+ architecture with optimal hyperparameters identified in Table 3. Red solid borders are physician (ground‐truth) beam apertures, and yellow dashed borders are predicted apertures. AP, anterior‐posterior; PA, posterior‐anterior; RL, right‐lateral field angle.