Search

PurposeTreatment plans are used for patients under radiotherapy in clinics. Before execution, these plans are checked for safety and quality by human experts. A few of them were identified with flaws and needed further improvement. To automate this checking process, an unsupervised learning method based on an autoencoder was proposed.MethodsFirst, features were extracted from the treatment plan by human experts. Then, these features were assembled and used for model learning. After network optimization, a reconstruction error between the predicted and target signals was obtained. Finally, the questionable plans were identified based on the value of the reconstruction error. A large value of the reconstruction error indicates a longer distance from the standard distribution of normal plans. A total of 576 treatment plans for breast cancer patients were used for the test. Among them, 19 were questionable plans identified by human experts. To evaluate the performance of the autoencoder, it was compared with four baseline detection algorithms, namely, local outlier factor (LOF), hierarchical density-based spatial clustering of applications with noise (HDBSCAN), one-class support vector machine (OC-SVM), and principal component analysis (PCA).ResultsThe results showed that the autoencoder achieved the best performance than the other four baseline algorithms. The AUC value of the autoencoder was 0.9985, while the second one was 0.9535 (LOF). While maintaining 100% recall, the average accuracy and precision of the results by the autoencoder were 0.9658 and 0.5143, respectively. While maintaining 100% recall, the average accuracy and precision of the results by LOF were 0.8090 and 0.1472, respectively.ConclusionThe autoencoder can effectively identify questionable plans from a large group of normal plans. There is no need to label the data and prepare the training data for model learning. The autoencoder provides an effective way to carry out an automatic plan checking in radiotherapy.

PurposeDeep-learning effectively predicts dose distributions in knowledge-based radiotherapy planning. Using anatomical information that includes a structure map and computed tomography (CT) data as input has been proven to work well. The minimum distance from each voxel in normal structures to planning target volume (DPTV) closely affects each voxel’s dose. In this study, we combined DPTV and anatomical information as input for a deep-learning–based dose-prediction network to improve performance.Materials and methodsOne hundred patients who underwent volumetric-modulated arc therapy for nasopharyngeal cancer were selected in this study. The prediction model based on a residual network had DPTV maps, structure maps, and CT as inputs and the corresponding dose distribution maps as outputs. The performances of the combined distance and anatomical information (COM) model and the traditional anatomical (ANAT) model with two-channel inputs (structure maps and CT) were compared. A 10-fold cross validation was performed to separately train and test the COM and ANAT models. The voxel-based mean error (ME), mean absolute error (MAE), dosimetric parameters, and dice similarity coefficient (DSC) of isodose volumes were used for modeling evaluation.ResultsThe mean MAE of the body volume of the COM model were 4.89 ± 1.35%, highly significantly lower than those for the ANAT model of 5.07 ± 1.37% (p<0.001). The ME values of the body for the 2-type models were similar (p >0.05). The mean DSC values of the isodose volumes in the range of 60 Gy were all better in the COM model (ppConclusionsThe COM model outperformed the ANAT model and could improve automated planning with statistically highly significant differences.

PurposeIn principle, intensity-modulated electron therapy (IMET) can be delivered through mechanical scanning, with a robotic arm mounting a linac.Materials and methodsHere is a scanning algorithm to identify the back-and-forth, top-to-bottom (zigzag) pattern scan sequence. The algorithm includes generating beam positions with a uniform resolution according to the applicator size; adopting discrete energies to achieve the depth of 90% dose by compositing energies; selecting energy by locating the target’s distal edge; and employing the energy-by-energy scan strategy for step-and-shoot discrete scanning. After a zigzag scan sequence is obtained, the delivery order of the scan spots is optimized by fast simulated annealing (FSA) to minimize the path length. For algorithm evaluation, scan sequences were generated using the computed tomography data of 10 patients with pancreatic cancer undergoing intraoperative radiotherapy, and the results were compared between the zigzag path and an optimized path. A simple calculation of the treatment delivery time, which comprises the irradiation time, the total robotic arm moving time, the time for energy switch, and the time to stop and restart the beam, was also made.ResultsIn these clinical cases, FSA optimization shortened the path lengths by 12%–43%. Assuming the prescribed dose was 15 Gy, machine dose rate was 15 Gy/s, energy switch time was 2 s, stop and restart beam time was 20 ms, and robotic arm move speed was 50 mm/s, the average delivery time was 124±38 s. The largest reduction in path length yielded an approximately 10% reduction in the delivery time, which can be further reduced by increasing the machine dose rate and the robotic arm speed, decreasing the time for energy switch, and/or developing more efficient algorithms.ConclusionMechanically scanning IMET is potentially feasible and worthy of further exploration.

PurposeThe challenge of cone-beam computed tomography (CBCT) is its low image quality, which limits its application for adaptive radiotherapy (ART). Despite recent substantial improvement in CBCT imaging using the deep learning method, the image quality still needs to be improved for effective ART application. Spurred by the advantages of transformers, which employs multi-head attention mechanisms to capture long-range contextual relations between image pixels, we proposed a novel transformer-based network (called TransCBCT) to generate synthetic CT (sCT) from CBCT. This study aimed to further improve the accuracy and efficiency of ART.Materials and methodsIn this study, 91 patients diagnosed with prostate cancer were enrolled. We constructed a transformer-based hierarchical encoder–decoder structure with skip connection, called TransCBCT. The network also employed several convolutional layers to capture local context. The proposed TransCBCT was trained and validated on 6,144 paired CBCT/deformed CT images from 76 patients and tested on 1,026 paired images from 15 patients. The performance of the proposed TransCBCT was compared with a widely recognized style transferring deep learning method, the cycle-consistent adversarial network (CycleGAN). We evaluated the image quality and clinical value (application in auto-segmentation and dose calculation) for ART need.ResultsTransCBCT had superior performance in generating sCT from CBCT. The mean absolute error of TransCBCT was 28.8 ± 16.7 HU, compared to 66.5 ± 13.2 for raw CBCT, and 34.3 ± 17.3 for CycleGAN. It can preserve the structure of raw CBCT and reduce artifacts. When applied in auto-segmentation, the Dice similarity coefficients of bladder and rectum between auto-segmentation and oncologist manual contours were 0.92 and 0.84 for TransCBCT, respectively, compared to 0.90 and 0.83 for CycleGAN. When applied in dose calculation, the gamma passing rate (1%/1 mm criterion) was 97.5% ± 1.1% for TransCBCT, compared to 96.9% ± 1.8% for CycleGAN.ConclusionsThe proposed TransCBCT can effectively generate sCT for CBCT. It has the potential to improve radiotherapy accuracy.

BackgroundRadical radiotherapy is the main treatment modality for early and locally advanced nasopharyngeal carcinoma (NPC). Magnetic resonance imaging (MRI) has the advantages of no ionizing radiation and high soft-tissue resolution compared to computed tomography (CT), but it does not provide electron density (ED) information for radiotherapy planning. Therefore, in this study, we developed a pseudo-CT (pCT) generation method to provide necessary ED information for MRI-only planning in NPC radiotherapy.MethodsTwenty patients with early-stage NPC who received radiotherapy in our hospital were investigated. First, 1433 sets of paired T1 weighted magnetic resonance (MR) simulation images and CT simulation images were rigidly registered and preprocessed. A 16-layer U-Net was used to train the pCT generative model and a “pix2pix” generative adversarial network (GAN) was also trained to compare with the pure U-Net regrading pCT quality. Second, the contours of all target volumes and organs at risk in the original CT were transferred to the pCT for planning, and the beams were copied back to the original CT for reference dose calculation. Finally, the dose distribution calculated on the pCT was compared with the reference dose distribution through gamma analysis and dose-volume indices.ResultsThe average time for pCT generation for each patient was 7.90 ± 0.47 seconds. The average mean (absolute) error was −9.3 ± 16.9 HU (102.6 ± 11.4 HU), and the mean-root-square error was 209.8 ± 22.6 HU. There was no significant difference between the pCT quality of pix2pix GAN and that of pure U-Net (p > 0.05). The dose distribution on the pCT was highly consistent with that on the original CT. The mean gamma pass rate (2 mm/3%, 10% low dose threshold) was 99.1% ± 0.3%, and the mean absolute difference of nasopharyngeal PGTV D99% and PTV V95% were 0.4% ± 0.2% and 0.1% ± 0.1%.ConclusionThe proposed deep learning model can accurately predict CT from MRI, and the generated pCT can be employed in precise dose calculations. It is of great significance to realize MRI-only planning in NPC radiotherapy, which can improve structure delineation and considerably reduce additional imaging dose, especially when an MR-guided linear accelerator is adopted for treatment.

Purpose: To investigate whether a deep learning-assisted contour (DLAC) could provide greater accuracy, inter-observer consistency, and efficiency compared with a manual contour (MC) of the clinical target volume (CTV) for non-small cell lung cancer (NSCLC) receiving postoperative radiotherapy (PORT).Materials and Methods: A deep dilated residual network was used to achieve the effective automatic contour of the CTV. Eleven junior physicians contoured CTVs on 19 patients by using both MC and DLAC methods independently. Compared with the ground truth, the accuracy of the contour was evaluated by using the Dice coefficient and mean distance to agreement (MDTA). The coefficient of variation (CV) and standard distance deviation (SDD) were rendered to measure the inter-observer variability or consistency. The time consumed for each of the two contouring methods was also compared.Results: A total of 418 CTV sets were generated. DLAC improved contour accuracy when compared with MC and was associated with a larger Dice coefficient (mean ± SD: 0.75 ± 0.06 vs. 0.72 ± 0.07, p < 0.001) and smaller MDTA (mean ± SD: 2.97 ± 0.91 mm vs. 3.07 ± 0.98 mm, p < 0.001). The DLAC was also associated with decreased inter-observer variability, with a smaller CV (mean ± SD: 0.129 ± 0.040 vs. 0.183 ± 0.043, p < 0.001) and SDD (mean ± SD: 0.47 ± 0.22 mm vs. 0.72 ± 0.41 mm, p < 0.001). In addition, a value of 35% of time saving was provided by the DLAC (median: 14.81 min vs. 9.59 min, p < 0.001).Conclusions: Compared with MC, the DLAC is a promising strategy to obtain superior accuracy, consistency, and efficiency for the PORT-CTV in NSCLC.

The purpose of this study was to evaluate the outcomes and toxicity experienced by cervical cancer patients with positive lymph nodes (LNs) who were treated with intensity-modulated radiation therapy (IMRT) and intracavitary brachytherapy (ICBT) plus concurrent chemotherapy. We retrospectively evaluated 108 cervical cancer patients with computed tomography (CT)-based positive LNs treated with IMRT and ICBT plus concurrent chemotherapy between 2009 and 2011. IMRT plans were designed to deliver 50 Gy to 95% of the planning target volume (PTV; cervical tumor, pelvis, and parametrium), with daily doses of 1.6–1.8 and 60–70 Gy to 95% of the planning gross tumor volume (PGTV)-LN (pelvic or para-aortic LNs), with daily doses of 2.0–2.2 Gy. Overall survival (OS) and progression-free survival (PFS) Kaplan–Meier curves were plotted. Acute and late toxicities were evaluated according to the Radiation Therapy Oncology Group and European Organization for Research and Treatment of Cancer toxicity criteria. Of the 108 cases, 45 were stage IIB and 63 were stage IIIB. The median follow-up was 65 months (range 2–83). Overall, the 5 year cumulative incidences of pelvic failure alone, distant failure alone, and synchronous pelvic and distant failure were 8.3, 12.9, and 8.3%, respectively. The 5 year OS rate was 67.6%, and the 5 year PFS rate was 53.7%. The 5 year cumulative incidence was 9.2% for late gastrointestinal and genitourinary toxicities of Grade ≥3 and 51.8% for acute leukopenia of Grade ≥3. The clinical results suggest that IMRT and ICBT with concurrent chemotherapy is an effective treatment, with acceptable toxicity, for advanced cervical cancer involving positive LNs.

Background: Radiotherapy is one of the main treatment methods for nasopharyngeal carcinoma (NPC). It requires exact delineation of the nasopharynx gross tumor volume (GTVnx), the metastatic lymph node gross tumor volume (GTVnd), the clinical target volume (CTV), and organs at risk (OARs) in the planning computed tomography (CT) images. However, this task is time-consuming and operator-dependent. In the present study, we developed an end-to-end deep deconvolutional neural network (DDNN) for segmentation of these targets. Methods: The proposed DDNN is an end-to-end architecture enabling fast training and testing. It consists of two important components: an encoder network and a decoder network. The encoder network was used to extract the visual features of a medical image and the decoder network was used to recover the original resolution by deploying deconvolution. A total of 230 patients diagnosed with NPC stage I or stage II were included in this study. Data from 184 patients were chosen randomly as a training set to adjust the parameters of DDNN, and the remaining 46 patients were the test set to assess the performance of the model. The Dice similarity coefficient (DSC) was used to quantify the segmentation results of the GTVnx, GTVnd, and CTV. In addition, the performance of DDNN was compared with the VGG-16 model. Results: The proposed DDNN method outperformed the VGG-16 in all the segmentation. The mean DSC values of DDNN were 80.9% for GTVnx, 62.3% for the GTVnd, and 82.6% for CTV, whereas VGG-16 obtained 72.3%, 33.7%, and 73.7% for the DSC values, respectively. Conclusions: DDNN can be used to segment the GTVnx and CTV accurately. The accuracy for the GTVnd segmentation was relatively low due to the considerable differences in its shape, volume, and location among patients. The accuracy is expected to increase with more training data and combination of MR images. In conclusion, DDNN has the potential to improve the consistency of contouring and streamline radiotherapy workflows, but careful human review and a considerable amount of editing will be required.