Improved accuracy of relative electron density and proton stopping power ratio through CycleGAN machine learning

Objective. Kilovoltage computed tomography (kVCT) is the cornerstone of radiotherapy treatment planning for delineating tissues and towards dose calculation. For the former, kVCT provides excellent contrast and signal-to-noise ratio. For the latter, kVCT may have greater uncertainty in determining relative electron density ( ρ e ) and proton stopping power ratio (SPR). Conversely, megavoltage CT (MVCT) may result in superior dose calculation accuracy. The purpose of this work was to convert kVCT HU to MVCT HU using deep learning to obtain higher accuracy ρ e and SPR. Approach. Tissue-mimicking phantoms were created to compare kVCT- and MVCT-determined ρ e and SPR to physical measurements. Using 100 head-and-neck datasets, an unpaired deep learning model was trained to learn the relationship between kVCTs and MVCTs, creating synthetic MVCTs (sMVCTs). Similarity metrics were calculated between kVCTs, sMVCTs, and MVCTs in 20 test datasets. An anthropomorphic head phantom containing bone-mimicking material with known composition was scanned to provide an independent determination of ρ e and SPR accuracy by sMVCT. Main results. In tissue-mimicking bone, ρ e errors were 2.20% versus 0.19% and SPR errors were 4.38% versus 0.22%, for kVCT versus MVCT, respectively. Compared to MVCT, in vivo mean difference (MD) values were 11 and 327 HU for kVCT and 2 and 3 HU for sMVCT in soft tissue and bone, respectively. ρ e MD decreased from 1.3% to 0.35% in soft tissue and 2.9% to 0.13% in bone, for kVCT and sMVCT, respectively. SPR MD decreased from 1.8% to 0.24% in soft tissue and 6.8% to 0.16% in bone, for kVCT and sMVCT, respectively. Relative to physical measurements, ρ e and SPR error in anthropomorphic bone decreased from 7.50% and 7.48% for kVCT to <1% for both MVCT and sMVCT. Significance. Deep learning can be used to map kVCT to sMVCT, suggesting higher accuracy ρ e and SPR is achievable with sMVCT versus kVCT.

Improved contrast and noise of megavoltage computed tomography (MVCT) through cycle-consistent generative machine learning

PURPOSE

Megavoltage computed tomography (MVCT) has been implemented on many radiation therapy treatment machines as a tomographic imaging modality that allows for three-dimensional visualization and localization of patient anatomy. Yet MVCT images exhibit lower contrast and greater noise than its kilovoltage CT (kVCT) counterpart. In this work, we sought to improve these disadvantages of MVCT images through an image-to-image-based machine learning transformation of MVCT and kVCT images. We demonstrated that by learning the style of kVCT images, MVCT images can be converted into high-quality synthetic kVCT (skVCT) images with higher contrast and lower noise, when compared to the original MVCT.

METHODS

Kilovoltage CT and MVCT images of 120 head and neck (H&N) cancer patients treated on an Accuray TomoHD system were retrospectively analyzed in this study. A cycle-consistent generative adversarial network (CycleGAN) machine learning, a variant of the generative adversarial network (GAN), was used to learn Hounsfield Unit (HU) transformations from MVCT to kVCT images, creating skVCT images. A formal mathematical proof is given describing the interplay between function sensitivity and input noise and how it applies to the error variance of a high-capacity function trained with noisy input data. Finally, we show how skVCT shares distributional similarity to kVCT for various macro-structures found in the body.

RESULTS

Signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were improved in skVCT images relative to the original MVCT images and were consistent with kVCT images. Specifically, skVCT CNR for muscle-fat, bone-fat, and bone-muscle improved to 14.8 ± 0.4, 122.7 ± 22.6, and 107.9 ± 22.4 compared with 1.6 ± 0.3, 7.6 ± 1.9, and 6.0 ± 1.7, respectively, in the original MVCT images and was more consistent with kVCT CNR values of 15.2 ± 0.8, 124.9 ± 27.0, and 109.7 ± 26.5, respectively. Noise was significantly reduced in skVCT images with SNR values improving by roughly an order of magnitude and consistent with kVCT SNR values. Axial slice mean (S-ME) and mean absolute error (S-MAE) agreement between kVCT and MVCT/skVCT improved, on average, from -16.0 and 109.1 HU to 8.4 and 76.9 HU, respectively.

CONCLUSIONS

A kVCT-like qualitative aid was generated from input MVCT data through a CycleGAN instance. This qualitative aid, skVCT, was robust toward embedded metallic material, dramatically improves HU alignment from MVCT, and appears perceptually similar to kVCT with SNR and CNR values equivalent to that of kVCT images.

DoseGAN: a generative adversarial network for synthetic dose prediction using attention-gated discrimination and generation

Deep learning algorithms have recently been developed that utilize patient anatomy and raw imaging information to predict radiation dose, as a means to increase treatment planning efficiency and improve radiotherapy plan quality. Current state-of-the-art techniques rely on convolutional neural networks (CNNs) that use pixel-to-pixel loss to update network parameters. However, stereotactic body radiotherapy (SBRT) dose is often heterogeneous, making it difficult to model using pixel-level loss. Generative adversarial networks (GANs) utilize adversarial learning that incorporates image-level loss and is better suited to learn from heterogeneous labels. However, GANs are difficult to train and rely on compromised architectures to facilitate convergence. This study suggests an attention-gated generative adversarial network (DoseGAN) to improve learning, increase model complexity, and reduce network redundancy by focusing on relevant anatomy. DoseGAN was compared to alternative state-of-the-art dose prediction algorithms using heterogeneity index, conformity index, and various dosimetric parameters. All algorithms were trained, validated, and tested using 141 prostate SBRT patients. DoseGAN was able to predict more realistic volumetric dosimetry compared to all other algorithms and achieved statistically significant improvement compared to all alternative algorithms for the V100 and V120 of the PTV, V60 of the rectum, and heterogeneity index.

Attention-Aware Discrimination for MR-to-CT Image Translation Using Cycle-Consistent Generative Adversarial Networks

Purpose

To suggest an attention-aware, cycle-consistent generative adversarial network (A-CycleGAN) enhanced with variational autoencoding (VAE) as a superior alternative to current state-of-the-art MR-to-CT image translation methods.

Materials and Methods

An attention-gating mechanism is incorporated into a discriminator network to encourage a more parsimonious use of network parameters, whereas VAE enhancement enables deeper discrimination architectures without inhibiting model convergence. Findings from 60 patients with head, neck, and brain cancer were used to train and validate A-CycleGAN, and findings from 30 patients were used for the holdout test set and were used to report final evaluation metric results using mean absolute error (MAE) and peak signal-to-noise ratio (PSNR).

Results

A-CycleGAN achieved superior results compared with U-Net, a generative adversarial network (GAN), and a cycle-consistent GAN. The A-CycleGAN averages, 95% confidence intervals (CIs), and Wilcoxon signed-rank two-sided test statistics are shown for MAE (19.61 [95% CI: 18.83, 20.39], P = .0104), structure similarity index metric (0.778 [95% CI: 0.758, 0.798], P = .0495), and PSNR (62.35 [95% CI: 61.80, 62.90], P = .0571).

Conclusion

A-CycleGANs were a superior alternative to state-of-the-art MR-to-CT image translation methods.© RSNA, 2020.

A convolutional neural network algorithm for automatic segmentation of head and neck organs at risk using deep lifelong learning

PURPOSE

This study suggests a lifelong learning-based convolutional neural network (LL-CNN) algorithm as a superior alternative to single-task learning approaches for automatic segmentation of head and neck (OARs) organs at risk.

METHODS AND MATERIALS

Lifelong learning-based convolutional neural network was trained on twelve head and neck OARs simultaneously using a multitask learning framework. Once the weights of the shared network were established, the final multitask convolutional layer was replaced by a single-task convolutional layer. The single-task transfer learning network was trained on each OAR separately with early stoppage. The accuracy of LL-CNN was assessed based on Dice score and root-mean-square error (RMSE) compared to manually delineated contours set as the gold standard. LL-CNN was compared with 2D-UNet, 3D-UNet, a single-task CNN (ST-CNN), and a pure multitask CNN (MT-CNN). Training, validation, and testing followed Kaggle competition rules, where 160 patients were used for training, 20 were used for internal validation, and 20 in a separate test set were used to report final prediction accuracies.

RESULTS

On average contours generated with LL-CNN had higher Dice coefficients and lower RMSE than 2D-UNet, 3D-Unet, ST- CNN, and MT-CNN. LL-CNN required ~72 hrs to train using a distributed learning framework on 2 Nvidia 1080Ti graphics processing units. LL-CNN required 20 s to predict all 12 OARs, which was approximately as fast as the fastest alternative methods with the exception of MT-CNN.

CONCLUSIONS

This study demonstrated that for head and neck organs at risk, LL-CNN achieves a prediction accuracy superior to all alternative algorithms.

Optimizing beam models for dosimetric accuracy over a wide range of treatments

This work presents a systematic approach for testing a dose calculation algorithm over a variety of conditions designed to span the possible range of clinical treatment plans. Using this method, a TrueBeam STx machine with high definition multi-leaf collimators (MLCs) was commissioned in the RayStation treatment planning system (TPS). The initial model parameters values were determined by comparing TPS calculations with standard measured depth dose and profile curves. The MLC leaf offset calibration was determined by comparing measured and calculated field edges utilizing a wide range of MLC retracted and over-travel positions. The radial fluence was adjusted using profiles through both the center and corners of the largest field size, and through measurements of small fields that were located at highly off-axis positions. The flattening filter source was adjusted to improve the TPS agreement for the output of MLC-defined fields with much larger jaw openings. The MLC leaf transmission and leaf end parameters were adjusted to optimize the TPS agreement for highly modulated intensity-modulated radiotherapy (IMRT) plans. The final model was validated for simple open fields, multiple field configurations, the TG 119 C-shape target test, and a battery of clinical IMRT and volumetric-modulated arc therapy (VMAT) plans. The commissioning process detected potential dosimetric errors of over 10% and resulted in a final model that provided in general 3% dosimetric accuracy. This study demonstrates the importance of using a variety of conditions to adjust a beam model and provides an effective framework for achieving high dosimetric accuracy.

DoseNet: a volumetric dose prediction algorithm using 3D fully-convolutional neural networks

The goal of this study is to demonstrate the feasibility of a novel fully-convolutional volumetric dose prediction neural network (DoseNet) and test its performance on a cohort of prostate stereotactic body radiotherapy (SBRT) patients. DoseNet is suggested as a superior alternative to U-Net and fully connected distance map-based neural networks for non-coplanar SBRT prostate dose prediction. DoseNet utilizes 3D convolutional downsampling with corresponding 3D deconvolutional upsampling to preserve memory while simultaneously increasing the receptive field of the network. DoseNet was implemented on 2 Nvidia 1080 Ti graphics processing units and utilizes a 3 phase learning protocol to help achieve convergence and improve generalization. DoseNet was trained, validated, and tested with 151 patients following Kaggle completion rules. The dosimetric quality of DoseNet was evaluated by comparing the predicted dose distribution with the clinically approved delivered dose distribution in terms of conformity index, heterogeneity index, and various clinically relevant dosimetric parameters. The results indicate that the DoseNet algorithm is a superior alternative to U-Net and fully connected methods for prostate SBRT patients. DoseNet required ~50.1 h to train, and ~0.83 s to make a prediction on a 128  ×  128  ×  64 voxel image. In conclusion, DoseNet is capable of making accurate volumetric dose predictions for non-coplanar SBRT prostate patients, while simultaneously preserving computational efficiency.

A Deep Look Into the Future of Quantitative Imaging in Oncology: A Statement of Working Principles and Proposal for Change

The adoption of enterprise digital imaging, along with the development of quantitative imaging methods and the re-emergence of statistical learning, has opened the opportunity for more personalized cancer treatments through transformative data science research. In the last 5 years, accumulating evidence has indicated that noninvasive advanced imaging analytics (i.e., radiomics) can reveal key components of tumor phenotype for multiple lesions at multiple time points over the course of treatment. Many groups using homegrown software have extracted engineered and deep quantitative features on 3-dimensional medical images for better spatial and longitudinal understanding of tumor biology and for the prediction of diverse outcomes. These developments could augment patient stratification and prognostication, buttressing emerging targeted therapeutic approaches. Unfortunately, the rapid growth in popularity of this immature scientific discipline has resulted in many early publications that miss key information or use underpowered patient data sets, without production of generalizable results. Quantitative imaging research is complex, and key principles should be followed to realize its full potential. The fields of quantitative imaging and radiomics in particular require a renewed focus on optimal study design and reporting practices, standardization, interpretability, data sharing, and clinical trials. Standardization of image acquisition, feature calculation, and statistical analysis (i.e., machine learning) are required for the field to move forward. A new data-sharing paradigm enacted among open and diverse participants (medical institutions, vendors and associations) should be embraced for faster development and comprehensive clinical validation of imaging biomarkers. In this review and critique of the field, we propose working principles and fundamental changes to the current scientific approach, with the goal of high-impact research and development of actionable prediction models that will yield more meaningful applications of precision cancer medicine.

A continuous arc delivery optimization algorithm for CyberKnife m6

PURPOSE

This study aims to reduce the delivery time of CyberKnife m6 treatments by allowing for noncoplanar continuous arc delivery. To achieve this, a novel noncoplanar continuous arc delivery optimization algorithm was developed for the CyberKnife m6 treatment system (CyberArc-m6).

METHODS AND MATERIALS

CyberArc-m6 uses a five-step overarching strategy, in which an initial set of beam geometries is determined, the robotic delivery path is calculated, direct aperture optimization is conducted, intermediate MLC configurations are extracted, and the final beam weights are computed for the continuous arc radiation source model. This algorithm was implemented on five prostate and three brain patients, previously planned using a conventional step-and-shoot CyberKnife m6 delivery technique. The dosimetric quality of the CyberArc-m6 plans was assessed using locally confined mutual information (LCMI), conformity index (CI), heterogeneity index (HI), and a variety of common clinical dosimetric objectives.

RESULTS

Using conservative optimization tuning parameters, CyberArc-m6 plans were able to achieve an average CI difference of 0.036 ± 0.025, an average HI difference of 0.046 ± 0.038, and an average LCMI of 0.920 ± 0.030 compared with the original CyberKnife m6 plans. Including a 5 s per minute image alignment time and a 5-min setup time, conservative CyberArc-m6 plans achieved an average treatment delivery speed up of 1.545x ± 0.305x compared with step-and-shoot plans.

CONCLUSIONS

The CyberArc-m6 algorithm was able to achieve dosimetrically similar plans compared to their step-and-shoot CyberKnife m6 counterparts, while simultaneously reducing treatment delivery times.

Correcting TG 119 confidence limits

PURPOSE

Task Group 119 (TG-119) has been adopted for evaluating the adequacy of intensity-modulated radiation therapy (IMRT) commissioning and for establishing patient-specific IMRT quality assurance (QA) passing criteria in clinical practice. TG-119 establishes 95% confidence limits (CLs), which help clinics identify systematic IMRT QA errors and identify outliers. In TG-119, the 95% CLs are established by fitting the Gamma Γ analysis passing rate results to an assumed distribution, then calculating the limit in which 95% of the data fall. CLs for a given dataset will depend greatly on the type of distribution used, and those determined by following the TG-119 guidelines are only valid if the underlying data follows a Gaussian distribution. Gaussian distributions assume symmetry about the mean, which would imply the possibility of negative Γ analysis failing rates. This study demonstrates that the gamma distribution is a more reasonable assumption for establishing CLs than the Gaussian distribution used in TG-119. Thus, the gamma distribution is suggested as a replacement to the conventional Gaussian statistical model used in TG-119.

MATERIALS AND METHODS

The moments estimator (ME) for the gamma family is used to obtain the CLs of the failing rates for all Γ analysis criteria. To demonstrate the congruence of the gamma distribution, the root mean squared error and the CL values for the MEs of the gamma and the Gaussian families were compared.

RESULTS

In this study, the empirical 95% CLs generated using 302 plans represent the ground truth, which resulted in a 91.83% passing rate using 3%/3 mm error local criteria. The gamma distribution underestimates the 95% CL by 0.09%, while the Gaussian distribution overestimates the 95% CL by 4.12%.

CONCLUSIONS

Although IMRT QA equipment may vary between clinics, the mathematical formalism presented in this study applies to any combination of planning and delivery systems. This study has demonstrated that a gamma distribution should be favored over a Gaussian distribution when establishing CLs for IMRT QA.

Canny edge-based deformable image registration

This work focuses on developing a 2D Canny edge-based deformable image registration (Canny DIR) algorithm to register in vivo white light images taken at various time points. This method uses a sparse interpolation deformation algorithm to sparsely register regions of the image with strong edge information. A stability criterion is enforced which removes regions of edges that do not deform in a smooth uniform manner. Using a synthetic mouse surface ground truth model, the accuracy of the Canny DIR algorithm was evaluated under axial rotation in the presence of deformation. The accuracy was also tested using fluorescent dye injections, which were then used for gamma analysis to establish a second ground truth. The results indicate that the Canny DIR algorithm performs better than rigid registration, intensity corrected Demons, and distinctive features for all evaluation matrices and ground truth scenarios. In conclusion Canny DIR performs well in the presence of the unique lighting and shading variations associated with white-light-based image registration.

Design of a portable imager for near-infrared visualization of cutaneous wounds

A portable imager developed for real-time imaging of cutaneous wounds in research settings is described. The imager consists of a high-resolution near-infrared CCD camera capable of detecting both bioluminescence and fluorescence illuminated by an LED ring with a rotatable filter wheel. All external components are integrated into a compact camera attachment. The device is demonstrated to have competitive performance with a commercial animal imaging enclosure box setup in beam uniformity and sensitivity. Specifically, the device was used to visualize the bioluminescence associated with increased reactive oxygen species activity during the wound healing process in a cutaneous wound inflammation model. In addition, this device was employed to observe the fluorescence associated with the activity of matrix metalloproteinases in a mouse lipopolysaccharide-induced infection model. Our results support the use of the portable imager design as a noninvasive and real-time imaging tool to assess the extent of wound inflammation and infection.

An accurate algorithm to match imperfectly matched images for lung tumor detection without markers

In order to locate lung tumors on kV projection images without internal markers, digitally reconstructed radiographs (DRRs) are created and compared with projection images. However, lung tumors always move due to respiration and their locations change on projection images while they are static on DRRs. In addition, global image intensity discrepancies exist between DRRs and projections due to their different image orientations, scattering, and noises. This adversely affects comparison accuracy. A simple but efficient comparison algorithm is reported to match imperfectly matched projection images and DRRs. The kV projection images were matched with different DRRs in two steps. Preprocessing was performed in advance to generate two sets of DRRs. The tumors were removed from the planning 3D CT for a single phase of planning 4D CT images using planning contours of tumors. DRRs of background and DRRs of tumors were generated separately for every projection angle. The first step was to match projection images with DRRs of background signals. This method divided global images into a matrix of small tiles and similarities were evaluated by calculating normalized cross-correlation (NCC) between corresponding tiles on projections and DRRs. The tile configuration (tile locations) was automatically optimized to keep the tumor within a single projection tile that had a bad matching with the corresponding DRR tile. A pixel-based linear transformation was determined by linear interpolations of tile transformation results obtained during tile matching. The background DRRs were transformed to the projection image level and subtracted from it. The resulting subtracted image now contained only the tumor. The second step was to register DRRs of tumors to the subtracted image to locate the tumor. This method was successfully applied to kV fluoro images (about 1000 images) acquired on a Vero (BrainLAB) for dynamic tumor tracking on phantom studies. Radiation opaque markers were implanted and used as ground truth for tumor positions. Although other organs and bony structures introduced strong signals superimposed on tumors at some angles, this method accurately located tumors on every projection over 12 gantry angles. The maximum error was less than 2.2 mm, while the total average error was less than 0.9mm. This algorithm was capable of detecting tumors without markers, despite strong background signals.

SU-E-J-185: 4DCT Geometrical Eigenmode Model for Inter-Fraction Evaluation of Tumor Regression and Breathing Pattern Changes

PURPOSE

Evaluate the feasibility of using intra-fractional respiratory dominant eigenmodes of lung deformation as a method to track inter-fractional tumor regression and other physiological changes. Geometrical evaluation by dominant eigenmodes during free-breathing may be implemented as an IGART tool for assessing inter-fractional non-respiratory anatomical variation.

METHODS

Intra-fractional deformable image registration (DIR) is performed on 4DCT scans of lung cancer patients. Principal component analysis is conducted on a set of 10 deformation vector fields for each patient both with and without a lung mask. Leave-one-out cross-validation (LOOCV) is used to assess intra-fractional modeling error among a subset of respiratory phases. The first two dominant eigenmodes of PCA are compared across patients in axial, sagittal, and coronal slices.

RESULTS

Eigenvalues between the first and second dominant respiratory modes decay by a factor of ten for all patients. Intra-fractional LOOCV maximum error is on the order of 1mm inside the lung and 4-6mm for the whole thorax for all patients using the first two dominant eigenmodes. Eigenvalue decay is fastest for patients where diaphragm motion is relatively larger in magnitude. The first two eigenmodes generally demonstrate a strong but smooth superior-inferior motion component with a lesser lateral and anterior-posterior component, while subsequent less dominant eigenmodes produce more random directional eigenvector fields.

CONCLUSION

The two first principal components are consistent amongst patients. Inter-fractional eigenmode comparison may provide a means to track tumor regression and breathing pattern or physiological changes. We expect that patients with more dominant diaphragmatic respiration will allow for more delineated eigenmode classification. Clear dominant eigenmode delineation may allow subsequent less dominant eigenmodes to be inter-fractionally compared as a means to evaluate and compensate for tumor regression.

A comparison of two different shuttle run tests for the estimation of VO2max

The aims of this experiment were twofold. The first was to determine whether there was a significant difference between two types of 20-m shuttle run test used to estimate VO2max, these being the Canadian version (CT) and the European versione (ET). The second aim was to determine which of the two tests best estimated direct VO2 measurement in our laboratory. To accomplish the first aim, 500 schoolchildren aged 12 to 16 years were randomly chosen from schools within Tasmania to undertake the two tests within seven days of each other. On the day of testing the children were assigned to one of the two tests and had no knowledge as to which test was being undertaken. Half of the children underwent the CT test first while the other half undertook the ET test first. Seven days after the first test was completed the appropriate second test was undertaken. The instructions to each child centred around the necessity to complete as many shuttles as possible staying in time with a pre-recorded signal. A relationship between the two sets of shuttle run data indicated that there was a significant correlation between the ET and CT, r = 0.834 (p < 0.0001). A Student's "t" test revealed that when the estimates of VO2max were compared however, there was a significant difference between the two tests (p < 0.0001). The ET estimated (Mean +/- SEM) VO2max at 34.9 +/- 0.45 ml.kg-1.min-1 whereas the CT estimated VO2max at 43.3 +/- 0.40 ml.kg-1.min-1. When this data was correlated, the co-efficient dropped to r = 0.761 which was still significant (p < 0.001). In order to accomplish the second aim, fifty children were chosen at random to undertake a VO2max test (DM) which was conducted via standard open circuit spirometry using a Quinton Metabolic Cart (QMC). The highest correlation was DM:ET being r = 0.93 whereas DM:CT was r = 0.87, both being significant at p < 0.001. When the data was compared there was a significant (p < 0.05) difference between DM and ET. DM measured VO2max as 37.6 +/- 0.37 ml.kg-1.min-1 whereas ET underestimated DM and measured VO2max at 34.7 +/- 0.56 ml.kg-1.min-1. The CT (41.9 +/- 0.62 ml.kg-1.min-1) over estimated DM by 11.4% however, the difference here also being significant (p < 0.01). This results of this study would suggest that teachers and coaches should use either one test or the other in the estimation of VO2max as the two tests differ significantly in their estimation. Of these two test versions, the ET underestimates direct VO2max measurement but is more accurate than the CT, so we feel this is the test of choice.