Table 1.
Selected studies on deep learning-based automated radiotherapy planning.
| Reference | Year | Network | Training sets | Test sets/NO. | Input | Output | Results | Research Highlight |
|---|---|---|---|---|---|---|---|---|
| Shiraishi et al. (18) | 2016 | ANN | 23 prostate and 43 SRS/SRT VMAT plans. Twelve training and 11 validation for prostate, and 23 training and 20 validation for SRS/SRT | No | Manually determined geometric and plan parameters | 3D dose | Prediction errors <10% and 8% for prostate and SRS/SRT cases, respectively | Knowledge-based 3D dose predictions, rather than previous 1D DVH prediction |
| Campbell et al. (21) | 2017 | ANN | 43 pancreatic Arc-based SBRT patients. Nineteen training and 10 validation for Group A, 9 and 5 for Group B, respectively | No | Plan parameters and voxel-based geometric parameters | 3D dose | Mean dose error <5% | Prediction accuracy substantially improved when each physician's treatment approach was taken into account by training their own dedicated models |
| Nguyen et al. (22) | 2017 | Modified 2D-Unet | 80 prostate IMRT patients, 10-fold cross-validation | 8 | labeled targets and OARs | 3D dose | Prediction errors around 2% in PTVs and under 5% of the prescription dose in OARs, isodose volumes average dice coefficient of 0.91 | Unet for 3D dose prediction |
| Nguyen et al. (23) | 2019 | 3D HD U-Net | 100 H&N VMAT patients, 5-fold cross validation | 20 | Labeled targets and OARs, prescription doses | 3D Dose | OARs dose difference :maximum error within 6.3% and mean error within 5.1% | Outperforming the Standard U-net and Dense-Net in both prediction accuracy and efficiency |
| Barragán-Montero et al. (26) | 2019 | 3D HD U-Net | 100 lung IMRT patients, training, and validation | 29 | Labeled targets and OARs, beam setup information | 3D Dose | Prediction accuracy improved substantially in low and medium dose regions and slightly in high dose regions | Prediction accuracy improved by considering beam setup information |
| Zhou et al. (27) | 2020 | 3D U-Res-Net | 100 rectal cancer postoperative IMRT patients | 22 | Labeled targets and OARs, beam setup information | 3D Dose | Mean absolute prediction errors 3.92 ± 4.16%,clearly outperforming 3D U-Res-Net_O and slightly superior to 3D U-Net | Prediction accuracy improved by considering beam setup information |
| Kearney et al. (28) | 2018 | FCNN Dose-Net | 126 prostate non-coplanar SBRT Cyberknife patients, 106 training, 20 validation | 25 | Labeled targets and OARs, dose prescription | 3D Dose | A superior alternative to U-Net and fully connected network | Utilizes a 3 phase learning protocol to achieve convergence and improve generalization |
| Kajikawa et al. (29) | 2018 | Alex-Net CNN | 60 prostate IMRT patients, five-fold cross-validation | No | CT images, with/without labeled structures | 3D dose | Prediction accuracies 56.7 ± 9.7% and 70.0 ± 11.3%, respectively | Pre-trained on Image-Net database, the model with structure labels focused on areas related to dose constraints improved prediction accuracy |
| Chen et al. (30) | 2018 | Transfer learning ResNet | 70 early-stage NPC IMRTpatients | 10 | Labeled targets and OARs, with/without beam setup information | 2D dose map | Out-of-field dose distributions prediction error 4.7 ± 6.1%vs. 5.5 ±7.9%, input with/without beam setup information | Input information from beam geometry improved the out-of-field dose distributions prediction accuracy |
| Liu et al. (31) | 2019 | U-ResNet-D | 170 NPCTomotherapy patients, 136 training, 34 validation | 20 | Labeled targets and OARs,3D dose | 3D Dose | Mean absolute dose differences for PTVs and OARs are within 2.0 and 4.2%, respectively | U-ResNet-D for Tomotherapy 3D dose prediction |
| Fan et al. (32) | 2019 | ResNet | 270 H&N IMRT patients, 195 training, 25 validation | 50 | Labeled targets and OARs | 3D Dose | Predicted differences not statistically significant for clinical indices of all targets and OARs except the difference of 0.5% for PTV70.4 | Automatic plan generation based on predicted 3D dose distribution |
| Mahmood et al. (33) | 2018 | GAN | 130 oropharyngeal IMRT patients | 87 | Labeled targets and OARs, dose maps | 3D dose | Outperformed a query-based, a PCA-based method, a random forest, and a CNN method, on clinical criteria satisfaction | Recast the dose prediction problem as an image colorization problem, solve the problem using a GAN by mimicking the iterative process between the planner and oncologist |
| Appenzoller et al. (8) | 2019 | 3D CNN | 80 prostate IMRT patients | 15 | Labeled targets and OARs | 3D dose | Prediction error: 1.10 ± 0.64%, 2.50 ± 1.17%, 2.04 ± 1.40, and 2.08 ± 1.99% for D2,D98 in PTV-1 and V65 in rectum and V65 in bladder | 3D CNN was superior to or comparable with RapidPlanTM |
| Krayenbuehl et al. (50) | 2019 | CNN | 60 prostate VMAT patients | 10 | Labeled targets and OARs, the dose distribution from a PTV-only plan | 3D dose | Mean SARs for the PTV, bladder, and rectum 0.007 ± 0.003, 0.035 ± 0.032, and 0.067 ± 0.037, respectively | Prediction results better than the contours-based method |
| Shin et al. (53) | 2019 | DNN | 240 prostate IMRT plans | 45 | Labeled targets and OARs, dose distributions | Fluence-maps | Generated plan qualities comparable with the corresponding clinical plans | Generate beam fluence—maps directly from the organ contours and dose distributions without inverse planning |
| Wieser et al. (54) | 2020 | DNN,DRL-based VTPN | 10 prostate IMRT patients | 64 | IMRT plans | IMRT plans | Spontaneously learn how to adjust treatment planning parameters, high-quality treatment plans generated | The first artificial intelligence system to model the behaviors of human planners in treatment planning |
ANN, artificial neural network; SRS, stereotactic radiosurgery; SRT, stereotactic radiotherapy; VMAT, volumetric-modulated arc therapy; 3D, three dimensional; 1D, one dimensional; DVH, dose-volume histogram; SBRT, stereotactic body radiation therapy; 2D, two dimensional; IMRT, intensity-modulated radiation therapy; OARs, organs at risk; PTV, planning target volume; HD U-Net, Hierarchically densely connected U-Net; H&N, head and neck; U-ResNet-D, model looks like U-net, but uses ResNet to do down-sampling and deconvolution to perform up-sampling; FCNN, fully convolutional neural network; NPC, nasopharynx cancer; GAN, generative adversarial network; PCA, principal component analysis; MAE, mean absolute errors; SARs, sum of absolute residuals; DNN, deep-neural-network; DRL, deep reinforcement learning; VTPN, virtual treatment planner network.