Accurate segmentation of head and neck radiotherapy CT scans with 3D CNNs: consistency is key.

Objective.Automatic segmentation of organs-at-risk in radiotherapy planning computed tomography (CT) scans using convolutional neural networks (CNNs) is an active research area. Very large datasets are usually required to train such CNN models. In radiotherapy, large, high-quality datasets are scarce and combining data from several sources can reduce the consistency of training segmentations. It is therefore important to understand the impact of training data quality on the performance of auto-segmentation models for radiotherapy.Approach.In this study, we took an existing 3D CNN architecture for head and neck CT auto-segmentation and compare the performance of models trained with a small, well-curated dataset (n= 34) and then a far larger dataset (n= 185) containing less consistent training segmentations. We performed 5-fold cross-validations in each dataset and tested segmentation performance using the 95th percentile Hausdorff distance and mean distance-to-agreement metrics. Finally, we validated the generalisability of our models with an external cohort of patient data (n= 12) with five expert annotators.Main results.The models trained with a large dataset were greatly outperformed by models (of identical architecture) trained with a smaller, but higher consistency set of training samples. Our models trained with a small dataset produce segmentations of similar accuracy as expert human observers and generalised well to new data, performing within inter-observer variation.Significance.We empirically demonstrate the importance of highly consistent training samples when training a 3D auto-segmentation model for use in radiotherapy. Crucially, it is the consistency of the training segmentations which had a greater impact on model performance rather than the size of the dataset used.

Optimising a 3D convolutional neural network for head and neck computed tomography segmentation with limited training data.

Background and purpose: Convolutional neural networks (CNNs) are increasingly used to automate segmentation for radiotherapy planning, where accurate segmentation of organs-at-risk (OARs) is crucial. Training CNNs often requires large amounts of data. However, large, high quality datasets are scarce. The aim of this study was to develop a CNN capable of accurate head and neck (HN) 3D auto-segmentation of planning CT scans using a small training dataset (34 CTs). Materials and Method: Elements of our custom CNN architecture were varied to optimise segmentation performance. We tested and evaluated the impact of: using multiple contrast channels for the CT scan input at specific soft tissue and bony anatomy windows, resize vs. transpose convolutions, and loss functions based on overlap metrics and cross-entropy in different combinations. Model segmentation performance was compared with the inter-observer deviation of two doctors' gold standard segmentations using the 95th percentile Hausdorff distance and mean distance-to-agreement (mDTA). The best performing configuration was further validated on a popular public dataset to compare with state-of-the-art (SOTA) auto-segmentation methods. Results: Our best performing CNN configuration was competitive with current SOTA methods when evaluated on the public dataset with mDTA of ( 0.81 ± 0.31 ) mm for the brainstem, ( 0.20 ± 0.08 ) mm for the mandible, ( 0.77 ± 0.14 ) mm for the left parotid and ( 0.81 ± 0.28 ) mm for the right parotid. Conclusions: Through careful tuning and customisation we trained a 3D CNN with a small dataset to produce segmentations of HN OARs with an accuracy that is comparable with inter-clinician deviations. Our proposed model performed competitively with current SOTA methods.

Predictive value of vascular calcification identified in 4D planning CT of lung cancer patients treated with stereotactic body radiation therapy.

PURPOSE: The aim was to identify vascular calcification in 4DCT scan of lung cancer patients and establish the association between overall survival (OS) and vascular calcification, as surrogate for vascular health. METHODS: Vascular calcification within the thoracic cavity were segmented in 334 lung cancer patients treated with stereotactic body radiation therapy (SBRT). This has been done automatically on 4D planning CT and average reconstruction scans. Correlation between cardiac comorbidity and calcification volumes was evaluated for patients with recorded Adult Co-Morbidity Evaluation (n = 303). Associations between the identified calcifications and OS were further investigated. RESULTS: The volume of calcification from the average scan was significantly lower than from each phase (p < 0.001). The highest level of correlations between cardiac comorbidity and volume of the calcifications were found for one phase representing inhale and two phases representing exhale with the least motion blurring due to respiration (p < 0.005). The volume of the calcifications was subsequently averaged over these three phases. The average of calcification volumes over the three phases (denoted by inhale-exhale) showed the highest likelihood in univariate analysis and was chosen as vascular calcification measure. Cox-model suggested that tumor volume (Hazard Ratio [HR] = 1.46, p < 0.01) and inhale-exhale volume (HR = 1.05, p < 0.05) are independent factors predicting OS after adjusting for age, sex, and performance status. CONCLUSION: It was feasible to use. It 4DCT scan for identifying thoracic calcifications in lung cancer patients treated with SBRT. Calcification volumes from inhale-exhale phases had the highest correlation with overall cardiac comorbidity and the average of the calcification volume obtained from these phases was an independent predictive factor for OS.

Improving anatomical mapping of complexly deformed anatomy for external beam radiotherapy and brachytherapy dose accumulation in cervical cancer.

PURPOSE: In the treatment of cervical cancer, large anatomical deformations, caused by, e.g., tumor shrinkage, bladder and rectum filling changes, organ sliding, and the presence of the brachytherapy (BT) applicator, prohibit the accumulation of external beam radiotherapy (EBRT) and BT dose distributions. This work proposes a structure-wise registration with vector field integration (SW+VF) to map the largely deformed anatomies between EBRT and BT, paving the way for 3D dose accumulation between EBRT and BT. METHODS: T2w-MRIs acquired before EBRT and as a part of the MRI-guided BT procedure for 12 cervical cancer patients, along with the manual delineations of the bladder, cervix-uterus, and rectum-sigmoid, were used for this study. A rigid transformation was used to align the bony anatomy in the MRIs. The proposed SW+VF method starts by automatically segmenting features in the area surrounding the delineated organs. Then, each organ and feature pair is registered independently using a feature-based nonrigid registration algorithm developed in-house. Additionally, a background transformation is calculated to account for areas far from all organs and features. In order to obtain one transformation that can be used for dose accumulation, the organ-based, feature-based, and the background transformations are combined into one vector field using a weighted sum, where the contribution of each transformation can be directly controlled by its extent of influence (scope size). The optimal scope sizes for organ-based and feature-based transformations were found by an exhaustive analysis. The anatomical correctness of the mapping was independently validated by measuring the residual distances after transformation for delineated structures inside the cervix-uterus (inner anatomical correctness), and for anatomical landmarks outside the organs in the surrounding region (outer anatomical correctness). The results of the proposed method were compared with the results of the rigid transformation and nonrigid registration of all structures together (AST). RESULTS: The rigid transformation achieved a good global alignment (mean outer anatomical correctness of 4.3 mm) but failed to align the deformed organs (mean inner anatomical correctness of 22.4 mm). Conversely, the AST registration produced a reasonable alignment for the organs (6.3 mm) but not for the surrounding region (16.9 mm). SW+VF registration achieved the best results for both regions (3.5 and 3.4 mm for the inner and outer anatomical correctness, respectively). All differences were significant (p < 0.02, Wilcoxon rank sum test). Additionally, optimization of the scope sizes determined that the method was robust for a large range of scope size values. CONCLUSIONS: The novel SW+VF method improved the mapping of large and complex deformations observed between EBRT and BT for cervical cancer patients. Future studies that quantify the mapping error in terms of dose errors are required to test the clinical applicability of dose accumulation by the SW+VF method.

Accurate CT∕MR vessel-guided nonrigid registration of largely deformed livers.

PURPOSE: Computer tomography (CT) scans are used for designing radiotherapy treatment plans. However, the tumor is often better visible in magnetic resonance (MR) images. For liver stereotactic body radiation therapy (SBRT), the planning CT scan is acquired while abdominal compression is applied to reduce tumor motion induced by breathing. However, diagnostic MR scans are acquired under voluntary breath-hold without the compression device. The resulting large differences in liver shape hinder the alignment of CT and MR image sets, which severely limits the integration of the information provided by these images. The purpose of the current study is to develop and validate a nonrigid registration method to align breath-hold MR images with abdominal-compressed CT images, using vessels that are automatically segmented within the liver. METHODS: Contrast-enhanced MR and CT images of seven patients with liver cancer were used for this study. The registration method combines automatic vessel segmentation with an adapted version of thin-plate spline robust point matching. The vessel segmentation uses a multiscale vesselness measure, which allows vessels of various thicknesses to be segmented. The nonrigid registration is point-based, and progressively improves the correspondence and transformation between two point sets. Moreover, the nonrigid registration is capable of identifying and handling outliers (points with no counterpart in the other set). We took advantage of the strengths of both methods and created a multiscale registration algorithm. First, thick vessels are registered, then with each new iteration thinner vessels are included in the registration (strategy A). We compared strategy A to a straightforward approach where vessels of various diameters are segmented and subsequently registered (strategy B). To assess the transformation accuracy, residual distances were calculated for vessel bifurcations. For anatomical validation, residual distances were calculated for additional anatomical landmarks within the liver. To estimate the extent of deformation, the residual distances for the aforementioned anatomical points were calculated after rigid registration. RESULTS: Liver deformations in the range of 2.8-10.7 mm were found after rigid registration of the CT and MR scans. Low residual distances for vessel bifurcations (average 1.6, range 1.3-1.9 mm) and additional anatomical landmarks (1.5, 1.1-2.4 mm) were found after nonrigid registration. A large amount of outliers were identified (25%-55%) caused by vessels present in only one of the image sets and false positives in the vesselness measure. The nonrigid registration was capable of handling these outliers as was demonstrated by the low residual distances. Both strategies yielded very similar results in registration accuracy, but strategy A was faster than strategy B (≥2.0 times). CONCLUSIONS: An accurate CT∕MR vessel-guided nonrigid registration for largely deformed livers was developed, tested, and validated. The method, combining vessel segmentation and point matching, was robust against differences in the segmented vessels. The authors conclude that nonrigid registration is required for accurate alignment of abdominal-compressed and uncompressed liver anatomy. Alignment of breath-hold MR and abdominal-compressed CT images can be used to improve tumor localization for liver SBRT.

Three-dimensional dose addition of external beam radiotherapy and brachytherapy for oropharyngeal patients using nonrigid registration.

PURPOSE: To develop and evaluate a method for adding dose distributions of combined external beam radiotherapy (EBRT) and brachytherapy (BT) for oropharyngeal patients. METHODS AND MATERIALS: Two computed tomography (CT) scans were used for 5 patients: the EBRT CT, used for EBRT planning, and the BT CT, acquired after catheter implantation. For each scan, the salivary glands and the chewing and swallowing muscles were contoured, and a dose distribution was calculated. A nonrigid transformation was obtained by registering the organs' surfaces. Then the BT dose distribution was mapped onto the EBRT dose distribution by applying the transformation obtained. To account for differences in fractionation, the physical doses were converted to equivalent dose in 2 Gy (EQD(2)), and the total dose was found by adding dose voxel by voxel. The robustness of the dose addition was investigated by varying delineations and input parameters of the registration method and by varying the α/ß parameter for EQD(2). The effect of the perturbations was quantified using dose-volume histograms (DVH) and gamma analyses (distance-to-agreement/dose-difference = 1 mm/1 Gy). RESULTS: The variations in input parameters and delineations caused only small perturbations in the DVH of the added dose distributions. For most organs the gamma index was low, and it was moderately elevated for organs lying in areas with a steep gradient (median gamma index ≤ 2.3 for constrictor muscles, ≤ 0.7 for all other organs). CONCLUSIONS: The presented method allows adding dose distributions of combined EBRT and BT for oropharyngeal patients. In general, the method is reliable and robust with respect to uncertainties in organ delineation, perturbations in input parameters of the method, and α/ß values.

A symmetric nonrigid registration method to handle large organ deformations in cervical cancer patients.

PURPOSE: Modern radiotherapy requires assessment of patient anatomical changes. By using unidirectional registration methods, the quantified anatomical changes are asymmetric, i.e., depend on the direction of the registration. Moreover, the registration is challenged by the large and complex organ deformations that can occur in, e.g., cervical cancer patients. The aim of this work was to develop, test, and validate a symmetric feature-based nonrigid registration method that can handle organs with large-scale deformations. METHODS: A symmetric version of the unidirectional thin plate spline robust point matching (TPS-RPM) algorithm was developed, implemented, tested, and validated. Tests were performed by using the delineated cervix and uterus and bladder in CT scans of five cervical cancer patients. For each patient, five CT scans with a large variability in organ shape, volume, and deformations were acquired. Both the symmetric and the unidirectional algorithm were employed to calculate the registration geometric accuracy (surface distance and surface coverage errors), the inverse consistency, the residual distances after transforming anatomical landmarks, and the registration time. Additionally, to facilitate the further use of our symmetric method, a large set of input parameters was tested. RESULTS: The developed symmetric algorithm handled successfully the registration of bladders with extreme volume change for which TPS-RPM failed. Compared to the unidirectional algorithm the symmetric algorithm improved, for the registration of organs with large volume change, the inverse consistency by 78% and the surface coverage by 46%. Similarly, for organs with small volume change, the symmetric algorithm improved the inverse consistency by 69% and the surface coverage by 13%. The method allowed for anatomically coherent registration in only 35 s for cervix-uterus and 151 s for bladder, while keeping the inverse consistency errors around 1 mm and the surface matching errors below 1 mm. Compared to rigid alignment the symmetric method reduced the residual distances between anatomical landmarks from a range of 5.8 +/- 2-70.1 +/- 20.1 mm to a range of 1.9 +/- 0.2-8.5 +/- 5.2 mm. CONCLUSIONS: The developed symmetric method could be employed to perform fast, accurate, consistent, and anatomically coherent registration of organs with large and complex deformations. Therefore, the method is a useful tool that could support further developments in high precision image guided radiotherapy.

A novel flexible framework with automatic feature correspondence optimization for nonrigid registration in radiotherapy.

Technical improvements in planning and dose delivery and in verification of patient positioning have substantially widened the therapeutic window for radiation treatment of cancer. However, changes in patient anatomy during the treatment limit the exploitation of these new techniques. To further improve radiation treatments, anatomical changes need to be modeled and accounted for Nonrigid registration can be used for this purpose. This article describes the design, the implementation, and the validation of a new framework for nonrigid registration for radiotherapy applications. The core of this framework is an improved version of the thin plate spline robust point matching (TPS-RPM) algorithm. The TPS-RPM algorithm estimates a global correspondence and a transformation between the points that represent organs of interest belonging to two image sets. However, the algorithm does not allow for the inclusion of prior knowledge on the correspondence of subset of points, and therefore, it can lead to inconsistent anatomical solutions. In this article TPS-RPM was improved by employing a novel correspondence filter that supports simultaneous registration of multiple structures. The improved method allows for coherent organ registration and for the inclusion of user-defined landmarks, lines, and surfaces inside and outside of structures of interest. A procedure to generate control points from segmented organs is described. The framework parameters r and lambda, which control the number of points and the nonrigidness of the transformation, respectively, were optimized for three sites with different degrees of deformation (head and neck, prostate, and cervix) using two cases per site. For the head and neck cases, the salivary glands were manually contoured on CT scans, for the prostate cases the prostate and the vesicles, and for the cervix cases the cervix uterus, the bladder, and the rectum. The transformation error obtained using the best set of parameters was below 1 mm for all the studied cases. The lengths of the deformation vectors were on average (+/- 1 standard deviation) 5.8 +/- 2.5 and 2.6 +/- 1.1 mm for the head and neck cases, 7.2 +/- 4.5 and 8.6 +/- 1.9 mm for the prostate cases, and 19.0 +/- 11.6 and 14.5 +/- 9.3 mm for the cervix cases. Distinguishable anatomical features were identified for each case and were used to validate the registration by calculating residual distances after transformation: 1.5 +/- 0.8, 2.3 +/- 1.0, and 6.3 +/- 2.9 mm for the head and neck, prostate, and cervix sites, respectively. Finally, the authors demonstrated how the inclusion of these anatomical features in the registration process reduced the residual distances to 0.8 +/- 0.5, 0.6 +/- 0.5, and 1.3 +/- 0.7 mm for the head and neck, prostate, and cervix sites, respectively. The inclusion of additional anatomical features produced more anatomically coherent transformations without compromising the transformation error. The authors concluded that the presented nonrigid registration framework is a powerful tool to simultaneously register multiple segmented organs with very different complexities.

Deformation of prostate and seminal vesicles relative to intraprostatic fiducial markers.

PURPOSE: To quantify the residual geometric uncertainties after on-line corrections with intraprostatic fiducial markers, this study analyzed the deformation of the prostate and, in particular, the seminal vesicles relative to such markers. PATIENTS AND METHODS: A planning computed tomography (CT) scan and three repeat CT scans were obtained for 21 prostate cancer patients who had had three to four cylindrical gold markers placed. The prostate and whole seminal vesicles (clinical target volume [CTV]) were delineated on each scan at a slice thickness of 1.5 mm. Rigid body transformations (translation and rotation) mapping the markers onto the planning scan positions were obtained. The translation only (T(only)) or both translation and rotation were applied to the delineated CTVs. Next, the residue CTV surface displacements were determined using nonrigid registration of the delineated contours. For translation and rotation of the CTV, the residues represented deformation; for T(only), the residues stemmed from deformation and rotation. T(only) represented the residues for most currently applied on-line protocols. The patient and population statistics of the CTV surface displacements were calculated. The intraobserver delineation variation was similarly quantified using repeat delineations for all patients and corrected for. RESULTS: The largest CTV deformations were observed at the anterior and posterior side of the seminal vesicles (population average standard deviation

Local anatomic changes in parotid and submandibular glands during radiotherapy for oropharynx cancer and correlation with dose, studied in detail with nonrigid registration.

PURPOSE: To quantify the anatomic changes caused by external beam radiotherapy in head-and-neck cancer patients in full three dimensions and to relate the local anatomic changes to the planned mean dose. METHODS AND MATERIALS: A nonrigid registration method was adapted for RT image registration. The method was applied in 10 head-and-neck cancer patients, who each underwent a planning and a repeat computed tomography scan. Contoured structures (parotid, submandibular glands, and tumor) were registered in a nonrigid manner. The accuracy of the transformation was determined. The transformation results were used to summarize the anatomic changes on a local scale for the irradiated and spared glands. The volume reduction of the glands was related to the planned mean dose. RESULTS: Transformation was accurate with a mean error of 0.6 +/- 0.5 mm. The volume of all glands and the primary tumor decreased. The lateral regions of the irradiated parotid glands moved inward (average, 3 mm), and the medial regions tended to remain in the same position. The irradiated submandibular glands shrank and moved upward. The spared glands showed only a small deformation ( approximately 1 mm in most regions). Overall, the primary tumors shrank. The volume loss of the parotid glands correlated significantly with the planned mean dose (p <0.001). CONCLUSION: General shrinkage and deformation of irradiated glands was seen. The spared glands showed few changes. These changes were assessed by a nonrigid registration method, which effectively described the local changes occurring in the head-and-neck region after external beam radiotherapy.

Target coverage in image-guided stereotactic body radiotherapy of liver tumors.

PURPOSE: To determine the effect of image-guided procedures (with computed tomography [CT] and electronic portal images before each treatment fraction) on target coverage in stereotactic body radiotherapy for liver patients using a stereotactic body frame (SBF) and abdominal compression. CT guidance was used to correct for day-to-day variations in the tumor's mean position in the SBF. METHODS AND MATERIALS: By retrospectively evaluating 57 treatment sessions, tumor coverage, as obtained with the clinically applied CT-guided protocol, was compared with that of alternative procedures. The internal target volume-plus (ITV(+)) was introduced to explicitly include uncertainties in tumor delineations resulting from CT-imaging artifacts caused by residual respiratory motion. Tumor coverage was defined as the volume overlap of the ITV(+), derived from a tumor delineated in a treatment CT scan, and the planning target volume. Patient stability in the SBF, after acquisition of the treatment CT scan, was evaluated by measuring the displacement of the bony anatomy in the electronic portal images relative to CT. RESULTS: Application of our clinical protocol (with setup corrections following from manual measurements of the distances between the contours of the planning target volume and the daily clinical target volume in three orthogonal planes, multiple two-dimensional) increased the frequency of nearly full (> or = 99%) ITV(+) coverage to 77% compared with 63% without setup correction. An automated three-dimensional method further improved the frequency to 96%. Patient displacements in the SBF were generally small (< or = 2 mm, 1 standard deviation), but large craniocaudal displacements (maximal 7.2 mm) were occasionally observed. CONCLUSION: Daily, CT-assisted patient setup may substantially improve tumor coverage, especially with the automated three-dimensional procedure. In the present treatment design, patient stability in the SBF should be verified with portal imaging.