A method of using deep learning to predict three‐dimensional dose distributions for intensity‐modulated radiotherapy of rectal cancer

Abstract Purpose To develop and test a three‐dimensional (3D) deep learning model for predicting 3D voxel‐wise dose distributions for intensity‐modulated radiotherapy (IMRT). Methods A total of 122 postoperative rectal cancer cases treated by IMRT were considered in the study, of which 100 cases were randomly selected as the training–validating set and the remaining as the testing set. A 3D deep learning model named 3D U‐Res‐Net_B was constructed to predict 3D dose distributions. Eight types of 3D matrices from CT images, contoured structures, and beam configurations were fed into the independent input channel, respectively, and the 3D matrix of dose distributions was taken as the output to train the 3D model. The obtained 3D model was used to predict new 3D dose distributions. The predicted accuracy was evaluated in two aspects: (a) The dice similarity coefficients (DSCs) of different isodose volumes, the average dose difference of all voxels within the body, and 3%/5 mm global gamma passing rates of organs at risks (OARs) and planned target volume (PTV) were used to address the spatial correspondence between predicted and clinical delivered 3D dose distributions; (b) The dosimetric index (DI) including homogeneity index, conformity index, V50, V45 for PTV and OARs between predicted and clinical truth were statistically analyzed with the paired‐samples t test. The model was also compared with 3D U‐Net and the same architecture model without beam configurations input (named as 3D U‐Res‐Net_O). Results The 3D U‐Res‐Net_B model predicted 3D dose distributions accurately. For the 22 testing cases, the average prediction bias ranged from −1.94% to 1.58%, and the overall mean absolute errors (MAEs) was 3.92 ± 4.16%; there was no statistically significant difference for nearly all DIs. The model had a DSCs value above 0.9 for most isodose volumes, and global 3D gamma passing rates varying from 0.81 to 0.90 for PTV and OARs, clearly outperforming 3D U‐Res‐Net_O and being slightly superior to 3D U‐Net. Conclusions This study developed a more general deep learning model by considering beam configurations input and achieved an accurate 3D voxel‐wise dose prediction for rectal cancer treated by IMRT, a potentially easier clinical implementation for more comprehensive automatic planning.

prediction for rectal cancer treated by IMRT, a potentially easier clinical implementation for more comprehensive automatic planning. When that situation occurs, the planner needs to manually add planning-auxiliary contours and reoptimize plan in order to refine the spatial dose distributions. Second, these methods usually rely on the extraction of handcrafted features, such as overlapping volume histogram, 21 In this paper, we proposed a 3D CNN model (called 3D U-Res-Net_B) based on 3D U-Net35 and Residual network36 to achieve voxel-wise dose prediction for postoperative rectal cancer patients treated by IMRT. First, eight types of 3D matrices were extracted from the CT images, contoured structures and beam configurations.
The matrices were then put into different input channels of the 3D U-Res-Net_B model, and the 3D matrix of clinically delivered dose distributions was used as the output. The 3D U-Res-Net_B can learn multiple-scale and multiple-level features of anatomy and beam configurations, and then map these features to 3D dose distributions.
In next sections, we introduce the patient data processing methods in Section 2.A, the 3D CNN model and training methods in Section 2.B, and the evaluation methods in Section 2.C. Then, we present experimental results in Section 3 and discuss the experimental results and related researches in Section 4.  (GE Healthcare, Waukesha, USA). The scanning range was from the lower edge of the L-2 vertebra to 5 cm below the ischial tubercle with a slice thickness of 5 mm. The images were reconstructed to 2.5 mm and transmitted to the Pinnacle 3 treatment planning system (Philips Radiation Oncology Systems, Fitchburg, WI, USA). The clinical target volume (CTV) and organs at risk (OARs) were delineated and checked by radiation oncologists, and a margin of 7 mm was applied around CTV to create the planned target volume (PTV) in consideration of the organ motion and positioning uncertainties. The IMRT plans were designed to deliver a prescription dose of 50 Gy to the PTV using 6-MV, five to seven coplanar beams in the "step and shoot" mode, and direct machine parameters optimization (DMPO) technique.

2.A | Patient database and data processing
The original CT images, contoured structures, beam configurations information, and delivery dose of the IMRT plan were exported from Pinnacle system and then converted to 3D matrices using a developed in-house python software program, as shown in  The Python deep learning library Keras37 with TensorFlow38 as backend was employed to achieve the 3D U-Res-Net_B architecture.
An Nvidia Geforce RTX 2080 GPU card with 8G memories was used to train the model. Once the model has been fully trained, it takes only a few seconds to predict 3D dose distribution for a new case.
An overview of the training and predicting process is illustrated in Fig. 3.

2.C | Prediction evaluation
To evaluate the performance of the proposed 3D U-Res-Net_B model, the 3D dose distributions, and DVH parameters of OARs and PTV were compared between the prediction and clinical truth.   were also calculated, where i stands for the voxel point and n is overall number of voxels within the body. Also, the performance of our model was evaluated using three commonly used metrics, including dice similarity coefficients (DSCs), Hausdorff distance 95% (HD 95 ), and mean surface distance (MSD). Dice similarity coefficients provides spatial overlap information, while HD 95 /MSD measure boundary similarity of different isodose surfaces between prediction and corresponding clinical truth, according to the following formulas, respectively.
where V iso-p and V iso-c denote the certain isodose volume of prediction and clinical truth, respectively, and S iso-p /S iso-c denote boundary sur- A global 3D gamma analysis, which is used as a tool for IMRT plan dose verification, was employed to further evaluate the accuracy of voxel-wise dose distribution prediction for OARs and PTV.
Dose difference and distance-to-agreement criterion were set to be 3% and 5 mm, respectively, and the gamma passing rates were calculated above the threshold of 5% prescription dose.
With respect to DVH parameters, first of all, the overall DVH where V ptv and V pres are the volume of PTV and the prescription dose region, respectively, and V ref is the irradiated PTV volume of the prescription dose.
In order to further evaluate the performance of our 3D U-Res-Net_B model, the model was compared with 3D U-Net35 on some performance metrics. In addition, the model was also compared with the 3D U-Res-Net_O model, not having the beam configurations input.

3.A.2 | Statistics of DVH dosimetric index
The overall DVH comparisons of PTV and OARs for four randomly selected testing patients between clinical and predicted results are presented in Fig. 6. The visual inspection indicate that the clinical and predicted DVHs of PTV and OARs have an acceptable agreement for each patient.

3.B | Comparison of model performance
The DSCs for 3D U-Res-Net_B, 3D U-Net and 3D U-Res-Net_O are presented in Fig. 7. The 3D U-Res-Net_B model has a DSCs value above 0.9 for most isodose volumes, clearly outperforming 3D U-Res-Net_O by 5% on average, and being slightly superior than 3D U-Net on average, with a value up to 3% higher. Additionally, a noticeable decline of DSCs at about 30 Gy isodose volumes is observed for all models, much most pronounced for 3D U-Res-Net_O model.      Fig. A1(a)], and the first combination of training-validation set (fold 1) for the aforemen-tioned model performance presentation and comparison. The obtained models were tested on the aforementioned testing set, and their own validation set, respectively. Figure A1(