Search

Background: Real‐time tumor tracking is one motion management method to address motion‐induced uncertainty. To date, fiducial markers are often required to reliably track lung tumors with X‐ray imaging, which carries risks of complications and leads to prolonged treatment time. A markerless tracking approach is thus desirable. Deep learning‐based approaches have shown promise for markerless tracking, but systematic evaluation and procedures to investigate applicability in individual cases are missing. Moreover, few efforts have been made to provide bounding box prediction and mask segmentation simultaneously, which could allow either rigid or deformable multi‐leaf collimator tracking. Purpose: The purpose of this study was to implement a deep learning‐based markerless lung tumor tracking model exploiting patient‐specific training which outputs both a bounding box and a mask segmentation simultaneously. We also aimed to compare the two kinds of predictions and to implement a specific procedure to understand the feasibility of markerless tracking on individual cases. Methods: We first trained a Retina U‐Net baseline model on digitally reconstructed radiographs (DRRs) generated from a public dataset containing 875 CT scans and corresponding lung nodule annotations. Afterwards, we used an independent cohort of 97 lung patients to develop a patient‐specific refinement procedure. In order to determine the optimal hyperparameters for automatic patient‐specific training, we selected 13 patients for validation where the baseline model predicted a bounding box on planning CT (PCT)‐DRR with intersection over union (IoU) with the ground‐truth higher than 0.7. The final test set contained the remaining 84 patients with varying PCT‐DRR IoU. For each testing patient, the baseline model was refined on the PCT‐DRR to generate a patient‐specific model, which was then tested on a separate 10‐phase 4DCT‐DRR to mimic the intrafraction motion during treatment. A template matching algorithm served as benchmark model. The testing results were evaluated by four metrics: the center of mass (COM) error and the Dice similarity coefficient (DSC) for segmentation masks, and the center of box (COB) error and the DSC for bounding box detections. Performance was compared to the benchmark model including statistical testing for significance. Results: A PCT‐DRR IoU value of 0.2 was shown to be the threshold dividing inconsistent (68%) and consistent (100%) success (defined as mean bounding box DSC > 0.6) of PS models on 4DCT‐DRRs. Thirty‐seven out of the eighty‐four testing cases had a PCT‐DRR IoU above 0.2. For these 37 cases, the mean COM error was 2.6 mm, the mean segmentation DSC was 0.78, the mean COB error was 2.7 mm, and the mean box DSC was 0.83. Including the validation cases, the model was applicable to 50 out of 97 patients when using the PCT‐DRR IoU threshold of 0.2. The inference time per frame was 170 ms. The model outperformed the benchmark model on all metrics, and the comparison was significant (p < 0.001) over the 37 PCT‐DRR IoU > 0.2 cases, but not over the undifferentiated 84 testing cases. Conclusions: The implemented patient‐specific refinement approach based on a pre‐trained baseline model was shown to be applicable to markerless tumor tracking in simulated radiographs for lung cases. [ABSTRACT FROM AUTHOR]

Deep learning models based on medical images play an increasingly important role for cancer outcome prediction. The standard approach involves usage of convolutional neural networks (CNNs) to automatically extract relevant features from the patient's image and perform a binary classification of the occurrence of a given clinical endpoint. In this work, a 2D-CNN and a 3D-CNN for the binary classification of distant metastasis (DM) occurrence in head and neck cancer patients were extended to perform time-to-event analysis. The newly built CNNs incorporate censoring information and output DM-free probability curves as a function of time for every patient. In total, 1037 patients were used to build and assess the performance of the time-to-event model. Training and validation was based on 294 patients also used in a previous benchmark classification study while for testing 743 patients from three independent cohorts were used. The best network could reproduce the good results from 3-fold cross validation [Harrell's concordance indices (HCIs) of 0.78, 0.74 and 0.80] in two out of three testing cohorts (HCIs of 0.88, 0.67 and 0.77). Additionally, the capability of the models for patient stratification into high and low-risk groups was investigated, the CNNs being able to significantly stratify all three testing cohorts. Results suggest that image-based deep learning models show good reliability for DM time-to-event analysis and could be used for treatment personalisation. [ABSTRACT FROM AUTHOR]

In presence of inter-fractional anatomical changes, clinical benefits are anticipated from image-guided adaptive radiotherapy. Nowadays, cone-beam CT (CBCT) imaging is mostly utilized during pre-treatment imaging for position verification. Due to various artifacts, image quality is typically not sufficient for photon or proton dose calculation, thus demanding accurate CBCT correction, as potentially provided by deep learning techniques. This work aimed at investigating the feasibility of utilizing a cycle-consistent generative adversarial network (cycleGAN) for prostate CBCT correction using unpaired training. Thirty-three patients were included. The network was trained to translate uncorrected, original CBCT images (CBCTorg) into planning CT equivalent images (CBCTcycleGAN). HU accuracy was determined by comparison to a previously validated CBCT correction technique (CBCTcor). Dosimetric accuracy was inferred for volumetric-modulated arc photon therapy (VMAT) and opposing single-field uniform dose (OSFUD) proton plans, optimized on CBCTcor and recalculated on CBCTcycleGAN. Single-sided SFUD proton plans were utilized to assess proton range accuracy. The mean HU error of CBCTcycleGAN with respect to CBCTcor decreased from 24 HU for CBCTorg to −6 HU. Dose calculation accuracy was high for VMAT, with average pass-rates of 100%/89% for a 2%/1% dose difference criterion. For proton OSFUD plans, the average pass-rate for a 2% dose difference criterion was 80%. Using a (2%, 2 mm) gamma criterion, the pass-rate was 96%. 93% of all analyzed SFUD profiles had a range agreement better than 3 mm. CBCT correction time was reduced from 6–10 min for CBCTcor to 10 s for CBCTcycleGAN. Our study demonstrated the feasibility of utilizing a cycleGAN for CBCT correction, achieving high dose calculation accuracy for VMAT. For proton therapy, further improvements may be required. Due to unpaired training, the approach does not rely on anatomically consistent training data or potentially inaccurate deformable image registration. The substantial speed-up for CBCT correction renders the method particularly interesting for adaptive radiotherapy. [ABSTRACT FROM AUTHOR]

Background: Current commercially available hybrid magnetic resonance linear accelerators (MR‐Linac) use 2D+t cine MR imaging to provide intra‐fractional motion monitoring. However, given the limited temporal resolution of cine MR imaging, target intra‐frame motion deterioration effects, resulting in effective time latency and motion artifacts in the image domain, can be appreciable, especially in the case of fast breathing. Purpose: The aim of this work is to investigate intra‐frame motion deterioration effects in MR‐guided radiotherapy (MRgRT) by simulating the motion‐corrupted image acquisition, and to explore the feasibility of deep‐learning‐based compensation approaches, relying on the intra‐frame motion information which is spatially and temporally encoded in the raw data (k‐space). Methods: An intra‐frame motion model was defined to simulate motion‐corrupted MR images, with 4D anthropomorphic digital phantoms being exploited to provide ground truth 2D+t cine MR sequences. A total number of 10 digital phantoms were generated for lung cancer patients, with randomly selected eight patients for training or validation and the remaining two for testing. The simulation code served as the data generator, and a dedicated motion pattern perturbation scheme was proposed to build the intra‐frame motion database, where three degrees of freedom were designed to guarantee the diversity of intra‐frame motion trajectories, enabling a thorough exploration in the domain of the potential anatomical structure positions. U‐Nets with three types of loss functions: L1 or L2 loss defined in image or Fourier domain, referred to as NNImgLoss‐L1, NNFloss‐L1 and NNL2‐Loss were trained to extract information from the motion‐corrupted image and used to estimate the ground truth final‐position image, corresponding to the end of the acquisition. Images before and after compensation were evaluated in terms of (i) image mean‐squared error (MSE) and mean absolute error (MAE), and (ii) accuracy of gross tumor volume (GTV) contouring, based on optical‐flow image registration. Results: Image degradation caused by intra‐frame motion was observed: for a linearly and fully acquired Cartesian readout k‐space trajectory, intra‐frame motion resulted in an imaging latency of approximately 50% of the acquisition time; in comparison, the motion artifacts exhibited only a negligible contribution to the overall geometric errors. All three compensation models led to a decrease in image MSE/MAE and GTV position offset compared to the motion‐corrupted image. In the investigated testing dataset for GTV contouring, the average dice similarity coefficients (DSC) improved from 88% to 96%, and the 95th percentile Hausdorff distance (HD95) dropped from 4.8 mm to 2.1 mm. Different models showed slight performance variations across different intra‐frame motion amplitude categories: NNImgLoss‐L1 excelled for small/medium amplitudes, whereas NNFloss‐L1 demonstrated higher DSC median values at larger amplitudes. The saliency maps of the motion‐corrupted image highlighted the major contribution of the later acquired k‐space data, as well as the edges of the moving anatomical structures at their final positions, during the model inference stage. Conclusions: Our results demonstrate the deep‐learning‐based approaches have the potential to compensate for intra‐frame motion by utilizing the later acquired data to drive the convergence of the earlier acquired k‐space components. [ABSTRACT FROM AUTHOR]

Objectives: Deep learning-based auto-segmentation of head and neck cancer (HNC) tumors is expected to have better reproducibility than manual delineation. Positron emission tomography (PET) and computed tomography (CT) are commonly used in tumor segmentation. However, current methods still face challenges in handling whole-body scans where a manual selection of a bounding box may be required. Moreover, different institutions might still apply different guidelines for tumor delineation. This study aimed at exploring the auto-localization and segmentation of HNC tumors from entire PET/CT scans and investigating the transferability of trained baseline models to external real world cohorts. Methods: We employed 2D Retina Unet to find HNC tumors from whole-body PET/CT and utilized a regular Unet to segment the union of the tumor and involved lymph nodes. In comparison, 2D/3D Retina Unets were also implemented to localize and segment the same target in an end-to-end manner. The segmentation performance was evaluated via Dice similarity coefficient (DSC) and Hausdorff distance 95th percentile (HD95). Delineated PET/CT scans from the HECKTOR challenge were used to train the baseline models by 5-fold cross-validation. Another 271 delineated PET/CTs from three different institutions (MAASTRO, CRO, BERLIN) were used for external testing. Finally, facility-specific transfer learning was applied to investigate the improvement of segmentation performance against baseline models. Results: Encouraging localization results were observed, achieving a maximum omnidirectional tumor center difference lower than 6.8 cm for external testing. The three baseline models yielded similar averaged cross-validation (CV) results with a DSC in a range of 0.71–0.75, while the averaged CV HD95 was 8.6, 10.7 and 9.8 mm for the regular Unet, 2D and 3D Retina Unets, respectively. More than a 10% drop in DSC and a 40% increase in HD95 were observed if the baseline models were tested on the three external cohorts directly. After the facility-specific training, an improvement in external testing was observed for all models. The regular Unet had the best DSC (0.70) for the MAASTRO cohort, and the best HD95 (7.8 and 7.9 mm) in the MAASTRO and CRO cohorts. The 2D Retina Unet had the best DSC (0.76 and 0.67) for the CRO and BERLIN cohorts, and the best HD95 (12.4 mm) for the BERLIN cohort. Conclusion: The regular Unet outperformed the other two baseline models in CV and most external testing cohorts. Facility-specific transfer learning can potentially improve HNC segmentation performance for individual institutions, where the 2D Retina Unets could achieve comparable or even better results than the regular Unet. [ABSTRACT FROM AUTHOR]

Background and purpose: Magnetic resonance imaging guided radiotherapy (MRgRT) offers treatment plan adaptation to the anatomy of the day. In the current MRgRT workflow, this requires the time consuming and repetitive task of manual delineation of organs-at-risk (OARs), which is also prone to inter- and intra-observer variability. Therefore, deep learning autosegmentation (DLAS) is becoming increasingly attractive. No investigation of its application to OARs in thoracic magnetic resonance images (MRIs) from MRgRT has been done so far. This study aimed to fill this gap. Materials and methods: 122 planning MRIs from patients treated at a 0.35 T MR-Linac were retrospectively collected. Using an 80/19/23 (training/validation/test) split, individual 3D U-Nets for segmentation of the left lung, right lung, heart, aorta, spinal canal and esophagus were trained. These were compared to the clinically used contours based on Dice similarity coefficient (DSC) and Hausdorff distance (HD). They were also graded on their clinical usability by a radiation oncologist. Results: Median DSC was 0.96, 0.96, 0.94, 0.90, 0.88 and 0.78 for left lung, right lung, heart, aorta, spinal canal and esophagus, respectively. Median 95th percentile values of the HD were 3.9, 5.3, 5.8, 3.0, 2.6 and 3.5 mm, respectively. The physician preferred the network generated contours over the clinical contours, deeming 85 out of 129 to not require any correction, 25 immediately usable for treatment planning, 15 requiring minor and 4 requiring major corrections. Conclusions: We trained 3D U-Nets on clinical MRI planning data which produced accurate delineations in the thoracic region. DLAS contours were preferred over the clinical contours. [ABSTRACT FROM AUTHOR]

Background: Online adaptive radiation therapy (RT) using hybrid magnetic resonance linear accelerators (MR-Linacs) can administer a tailored radiation dose at each treatment fraction. Daily MR imaging followed by organ and target segmentation adjustments allow to capture anatomical changes, improve target volume coverage, and reduce the risk of side effects. The introduction of automatic segmentation techniques could help to further improve the online adaptive workflow by shortening the re-contouring time and reducing intra- and inter-observer variability. In fractionated RT, prior knowledge, such as planning images and manual expert contours, is usually available before irradiation, but not used by current artificial intelligence-based autocontouring approaches. Purpose: The goal of this study was to train convolutional neural networks (CNNs) for automatic segmentation of bladder, rectum (organs at risk, OARs), and clinical target volume (CTV) for prostate cancer patients treated at 0.35 T MR-Linacs. Furthermore, we tested the CNNs generalization on data from independent facilities and compared them with the MR-Linac treatment planning system (TPS) propagated structures currently used in clinics. Finally, expert planning delineations were utilized for patient- (PS) and facility-specific (FS) transfer learning to improve auto-segmentation of CTV and OARs on fraction images. Methods: In this study, data from fractionated treatments at 0.35 T MR-Linacs were leveraged to develop a 3D U-Net-based automatic segmentation. Cohort C1 had 73 planning images and cohort C2 had 19 planning and 240 fraction images. The baseline models (BMs) were trained solely on C1 planning data using 53 MRIs for training and 10 for validation. To assess their accuracy, the models were tested on three data subsets: (i) 10 C1 planning images not used for training, (ii) 19 C2 planning,and (iii) 240 C2 fraction images.BMs also served as a starting point for FS and PS transfer learning, where the planning images from C2 were used for network parameter fine tuning.The segmentation output of the different trained models was compared against expert ground truth by means of geometric metrics.Moreover, a trained physician graded the network segmentations as well as the segmentations propagated by the clinical TPS. Results: The BMs showed dice similarity coefficients (DSC) of 0.88(4) and 0.93(3) for the rectum and the bladder, respectively, independent of the facility. CTV segmentation with the BM was the best for intermediate- and high-risk cancer patients from C1 with DSC=0.84(5) and worst for C2 with DSC=0.74(7). The PS transfer learning brought a significant improvement in the CTV segmentation, yielding DSC=0.72(4) for post-prostatectomy and low-risk patients and DSC=0.88(5) for intermediate- and high-risk patients. The FS training did not improve the segmentation accuracy considerably. The physician’s assessment of the TPS-propagated versus network-generated structures showed a clear advantage of the latter. Conclusions: The obtained results showed that the presented segmentation technique has potential to improve automatic segmentation for MR-guided RT. [ABSTRACT FROM AUTHOR]

Background: The evaluation of automatic segmentation algorithms is commonly performed using geometric metrics. An analysis based on dosimetric parameters might be more relevant in clinical practice but is often lacking in the literature. The aim of this study was to investigate the impact of state-of-the-art 3D U-Net-generated organ delineations on dose optimization in radiation therapy (RT) for prostate cancer patients.Methods: A database of 69 computed tomography images with prostate, bladder, and rectum delineations was used for single-label 3D U-Net training with dice similarity coefficient (DSC)-based loss. Volumetric modulated arc therapy (VMAT) plans have been generated for both manual and automatic segmentations with the same optimization settings. These were chosen to give consistent plans when applying perturbations to the manual segmentations. Contours were evaluated in terms of DSC, average and 95% Hausdorff distance (HD). Dose distributions were evaluated with the manual segmentation as reference using dose volume histogram (DVH) parameters and a 3%/3 mm gamma-criterion with 10% dose cut-off. A Pearson correlation coefficient between DSC and dosimetric metrics, i.e. gamma index and DVH parameters, has been calculated.Results: 3D U-Net-based segmentation achieved a DSC of 0.87 (0.03) for prostate, 0.97 (0.01) for bladder and 0.89 (0.04) for rectum. The mean and 95% HD were below 1.6 (0.4) and below 5 (4) mm, respectively. The DVH parameters, V[Formula: see text] for the bladder and V[Formula: see text] for the rectum, showed agreement between dose distributions within [Formula: see text] and [Formula: see text], respectively. The D[Formula: see text] and V[Formula: see text], for prostate and its 3 mm expansion (surrogate clinical target volume) showed agreement with the reference dose distribution within 2% and 3 Gy with the exception of one case. The average gamma pass-rate was 85%. The comparison between geometric and dosimetric metrics showed no strong statistically significant correlation.Conclusions: The 3D U-Net developed for this work achieved state-of-the-art geometrical performance. Analysis based on clinically relevant DVH parameters of VMAT plans demonstrated neither excessive dose increase to OARs nor substantial under/over-dosage of the target in all but one case. Yet the gamma analysis indicated several cases with low pass rates. The study highlighted the importance of adding dosimetric analysis to the standard geometric evaluation. [ABSTRACT FROM AUTHOR]

Purpose: To demonstrate a proof‐of‐concept for fast cone‐beam CT (CBCT) intensity correction in projection space by the use of deep learning. Methods: The CBCT scans and corresponding projections were acquired from 30 prostate cancer patients. Reference shading correction was performed using a validated method (CBCTcor), which estimates scatter and other low‐frequency deviations in the measured CBCT projections on the basis of a prior CT image obtained from warping the planning CT to the CBCT. A convolutional neural network (ScatterNet) was designed, consisting of an attenuation conversion stage followed by a shading correction stage using a UNet‐like architecture. The combined network was trained in 2D, utilizing pairs of measured and corrected projections of the reference method, in order to perform shading correction in projection space before reconstruction. The number of patients used for training, testing, and evaluation was 15, 7, and 8, respectively. The reconstructed CBCTScatterNet was compared to CBCTcor in terms of mean and absolute errors (ME and MAE) for the eight evaluation patients (not included in the network training). Volumetric modulated arc photon therapy (VMAT) and intensity‐modulated proton therapy (IMPT) plans were generated on CBCTcor. Dose was recalculated on CBCTScatterNet to evaluate its dosimetric accuracy. Single‐field uniform dose proton plans were utilized for proton range comparison of CBCTScatterNet and CBCTcor. Results: The CBCTScatterNet showed no cupping artifacts and a considerably smaller MAE and ME with respect to CBCTcor than the uncorrected CBCT (on average 144 Hounsfield units (HU) vs 46 HU for MAE and 138 HU vs −3 HU for ME). The pass‐rates using a 2% dose‐difference criterion at 50% dose cut‐off, were close to 100% for the VMAT plans of all patients when comparing CBCTScatterNet to CBCTcor. For IMPT plans pass‐rates were clearly lower, ranging from 15% to 81%. Proton range differences of up to 5 mm occurred. Conclusions: Using a deep convolutional neural network for CBCT intensity correction was shown to be feasible in the pelvic region for the first time. Dose calculation accuracy on CBCTScatterNet was high for VMAT, but unsatisfactory for IMPT. With respect to the reference technique (CBCTcor), the neural network enabled a considerable increase in speed for intensity correction and might eventually allow for on‐the‐fly shading correction during CBCT acquisition. [ABSTRACT FROM AUTHOR]

• Deep learning PET and CT-based DM and OS time-to-event models showed predictive capability. • PET-only achieved the best prognosis performance. • Unlike CT, GTV segmentation is not necessary for PET-based prognosis. Recent studies have shown that deep learning based on pre-treatment positron emission tomography (PET) or computed tomography (CT) is promising for distant metastasis (DM) and overall survival (OS) prognosis in head and neck cancer (HNC). However, lesion segmentation is typically required, resulting in a predictive power susceptible to variations in primary and lymph node gross tumor volume (GTV) segmentation. This study aimed at achieving prognosis without GTV segmentation, and extending single modality prognosis to joint PET/CT to allow investigating the predictive performance of combined- compared to single-modality inputs. We employed a 3D-Resnet combined with a time-to-event outcome model to incorporate censoring information. We focused on the prognosis of DM and OS for HNC patients. For each clinical endpoint, five models with PET and/or CT images as input were compared: PET-GTV, PET-only, CT-GTV, CT-only, and PET/CT-GTV models, where -GTV indicates that the corresponding images were masked using the GTV contour. Publicly available delineated CT and PET scans from 4 different Canadian hospitals (293) and the MAASTRO clinic (74) were used for training by 3-fold cross-validation (CV). For independent testing, we used 110 patients from a collaborating institution. The predictive performance was evaluated via Harrell's Concordance Index (HCI) and Kaplan-Meier curves. In a 5-year time-to-event analysis, all models could produce CV HCIs with median values around 0.8 for DM and 0.7 for OS. The best performance was obtained with the PET-only model, achieving a median testing HCI of 0.82 for DM and 0.69 for OS. Compared with the PET/CT-GTV model, the PET-only still had advantages of up to 0.07 in terms of testing HCI. The Kaplan-Meier curves and corresponding log-rank test results also demonstrated significant stratification capability of our models for the testing cohort. Deep learning-based DM and OS time-to-event models showed predictive capability and could provide indications for personalized RT. The best predictive performance achieved by the PET-only model suggested GTV segmentation might be less relevant for PET-based prognosis. [ABSTRACT FROM AUTHOR]