Feasibility of automated planning for whole‐brain radiation therapy using deep learning

Abstract Purpose The purpose of this study was to develop automated planning for whole‐brain radiation therapy (WBRT) using a U‐net‐based deep‐learning model for predicting the multileaf collimator (MLC) shape bypassing the contouring processes. Methods A dataset of 55 cases, including 40 training sets, five validation sets, and 10 test sets, was used to predict the static MLC shape. The digitally reconstructed radiograph (DRR) reconstructed from planning CT images as an input layer and the MLC shape as an output layer are connected one‐to‐one via the U‐net modeling. The Dice similarity coefficient (DSC) was used as the loss function in the training and ninefold cross‐validation. Dose‐volume‐histogram (DVH) curves were constructed for assessing the automatic MLC shaping performance. Deep‐learning (DL) and manually optimized (MO) approaches were compared based on the DVH curves and dose distributions. Results The ninefold cross‐validation ensemble test results were consistent with DSC values of 94.6 ± 0.4 and 94.7 ± 0.9 in training and validation learnings, respectively. The dose coverages of 95% target volume were (98.0 ± 0.7)% and (98.3 ± 0.8)%, and the maximum doses for the lens as critical organ‐at‐risk were 2.9 Gy and 3.9 Gy for DL and MO, respectively. The DL technique shows the consistent results in terms of the DVH parameter except for MLC shaping prediction for dose saving of small organs such as lens. Conclusions Comparable with the MO plan result, the WBRT plan quality obtained using the DL approach is clinically acceptable. Moreover, the DL approach enables WBRT auto‐planning without the time‐consuming manual MLC shaping and target contouring.


2.A | Datasets
Fifty-five patients, treated with WBRT in our hospital, were chosen.
They were treated with total doses of 30 Gy in 10-12 fractions. The treatment plans, which were designed in C-Linac 2300 iX and Vital-Beam (Varian Medical System, Palo, Alto, CA, USA) with a 120-leaf MLC, had two lateral radiation static fields (gantry angles of 90°and 270°) with the treatment isocenter at the center between the eyes.
The C-spine region was excluded from the treatment plans to prepare a consistent dataset to be localized in the brain region as a target. In those 55 cases, 45 cases were assigned to the training and validation sets, while the remaining 10 cases were assigned to the test set to evaluate the plan quality for unknown cases.  Fig. 1(a), was generated with Hounsfield unit range of (−100, 1000) from the CT image. The open-field region collimated by MLC was converted into a white color mask image, as shown in Fig. 1(b). The MLC shape [in Fig. 1(c)] predicted by the DL process was finally converted to the real MLC shape [in Fig. 1

2.B | U-net model and training
The DL framework, Tensorflow 18 with Keras, 19 was used for the automatic treatment planning of whole-brain cases. This analysis used the convolution neural network (CNN) architecture called "Unet,", 20 which is the convolutional encoder-decoder network widely used to prevent resolution loss for the image segmentation. The shapes of the DRR in the input layer and MLC in the output layer were one-to-one connected via the U-net modeling shown in Fig. 2.
Both the input and output layers consisted of 512 × 512 voxels with a single channel. The network for the downsampling of feature maps contained two-dimensional (2D) convolution (Conv2D) blocks and applied 3 × 3 filters with identical padding and rectified linear unit (ReLU) activation. Besides, max-pooling (MaxPooling2D) to reduce the pixel size and random deactivation of the unit (dropout) were employed. 21 For the first two convolution layer blocks, 50% dropout portions were achieved, and the dropout portion was 30% for the next two convolution layer blocks. The network for the upsampling of the feature map involved the inverse operation (Conv2DTranspose) of Conv2D and the concatenation process for better pixel localization regarding the input arrays. The output layer with the sigmoid activation was a 1 × 1 convolution with a single-channel converted from a previous layer with 32 channels. The number of network parameters was 7.7 M. An adaptive moment estimation (Adam) optimizer with a learning rate of 5 × 10 -5 was applied. Dice similarity coefficient (DSC) as loss function was used for the similarity assessment between the input and predicted MLC shapes. 22 The DSC is defined as −2|X i ∩ X p +S| / (|X i |∪|X p |+S), where X i (X p ) is an input (predicted) tensor obtained from image pixel values, and S is the smoothness (in this study, S = 1). The training was performed up

2.C | MLC shape prediction performance
The automatic MLC shaping using the DL technique was evaluated.
Dose-volume-histogram (DVH) curves constructed from the predicted MLC shape were compared with the MO results for dosimetric parameters of planning target volume (PTV) and organs-at-risk (OARs) such as lenses, eyes, and brainstem. The predicted MLC parameters on the 90°and 270°fields were imported to TPS. Dose calculations in given static MLC shaping were performed by the dose calculation algorithm of the analytical anisotropic algorithm (AAA) in Eclipse under the delivered monitor unit (MU) setting used in the previous treatment plan. 23 For 10 test cases, excluded from the training and validation dataset, the consistency and deviations of DVH curves were used to assess the learning performance. Also, the relative dose covering of PTV (D 95% , D 90% , D 50% , D 10% , D 5% , and D max ) and the absolute maximum dose (D max ) were calculated to estimate the target coverage and the OAR saving chances, respectively. Based on the statistical function library of SciPy v1.1.0 implemented on Python, the paired t-test with a two-tailed option was used to compare two data samples of MO and DL for DVH parameters. The significance level was set to P < 0.05 in this t-test.

| DISCUSSION
The feasibility of the automatic MLC shape prediction planning using the DL technique was evaluated in this study with whole-brain cases. To the best of our knowledge, this is the first study that predicts the MLC shape without additional contouring tasks using the U-net DL model based on 2D DRR reconstructed from CT images.
For the usual whole-brain planning, the shaping of MLC leaves does not follow the PTV shape, which means that the planning is  MLC shape prediction to small organs such as lenses should be also improved by the fine-tuned DL modeling and the larger dataset. To achieve complete automated planning, MU estimation for a given prescribed dose is another exciting topic.

| CONCLUSION
In this study, we evaluated training results using the DL technique with a dataset of 2D DRR images for whole-brain cases. The predicted plan quality was clinically acceptable based on the DVH curves and was comparable with the result of the MO plan. This DLimplemented planning without manual MLC shaping based on target contouring can help save time in the entire treatment process. It has the potential to improve the plan quality and enable rapid treatment in the whole-brain radiation therapy.