A methodology for evaluation of boundary detection algorithms on medical images

Assessment of the use of synthetic CT produced by deformable image registration of planning CT and CBCT in adaptive radiotherapy treatments of head and neck cancers

INTRODUCTION

The present study uses deformable image registration to produce synthetic CT (sCT) images and investigate their use in treatment planning and improving clinical judgment in assessing the need for adaptive radiotherapy (ART).

METHODS

A total of 30 patients with squamous cell carcinoma of the head and neck (HNSCC) who underwent ART were included in this study. The patients considered for adaptive planning, were re-simulated within 1-2 days. Both the Day 1 planning CT (pCT) and acquired CBCT were imported in Velocity® and a new sCT was created. The new treatment plan on re-simulation CT (rCT) which was planned for treatment delivery was recalculated on sCT. The geometric differences (Volume, Dice Similarity coefficient (DSC), mean distance to agreement (MDA)) in structures and dosimetric differences (Gamma analysis, mean doses and other DVH parameters) in treatment plans between the two images (sCT and rCT) were compared.

RESULTS

The evaluation between sCT and rCT revealed that the average DSC and MDA for all the structures obtained was 0.86(0.05) and 1.15(0.20) respectively. Global gamma passing rate for 3 %, 3 mm was 96.85 ± 2.10 %. Mean dose for OARs and PTVs were found to be similar (difference within 3 %) in the two images.

CONCLUSION

sCT can be used to predict the per fraction dose delivered to the patient and could be a better alternative than only relying on clinical judgments, to take the patient for ART. Further work needs to be done on the use of sCT images to replace rCT images for ART.

Efficient labeling for fine-tuning chest X-ray bone-suppression networks for pediatric patients

BACKGROUND

Pneumonia, a major infectious cause of morbidity and mortality among children worldwide, is typically diagnosed using low-dose pediatric chest X-ray [CXR (chest radiography)]. In pediatric CXR images, bone occlusion leads to a risk of missed diagnosis. Deep learning-based bone-suppression networks relying on training data have enabled considerable progress to be achieved in bone suppression in adult CXR images; however, these networks have poor generalizability to pediatric CXR images because of the lack of labeled pediatric CXR images (i.e., bone images vs. soft-tissue images). Dual-energy subtraction imaging approaches are capable of producing labeled adult CXR images; however, their application is limited because they require specialized equipment, and they are infrequently employed in pediatric settings. Traditional image processing-based models can be used to label pediatric CXR images, but they are semiautomatic and have suboptimal performance.

PURPOSE

We developed an efficient labeling approach for fine-tuning pediatric CXR bone-suppression networks capable of automatically suppressing bone structures in CXR images for pediatric patients without the need for specialized equipment and technologist training.

METHODS

Three steps were employed to label pediatric CXR images and fine-tune pediatric bone-suppression networks: distance transform-based bone-edge detection, traditional image processing-based bone suppression, and fully automated pediatric bone suppression. In distance transform-based bone-edge detection, bone edges were automatically detected by predicting bone-edge distance-transform images, which were then used as inputs in traditional image processing. In this processing, pediatric CXR images were labeled by obtaining bone images through a series of traditional image processing techniques. Finally, the pediatric bone-suppression network was fine-tuned using the labeled pediatric CXR images. This network was initially pretrained on a public adult dataset comprising 240 adult CXR images (A240) and then fine-tuned and validated on 40 pediatric CXR images (P260_40labeled) from our customized dataset (named P260) through five-fold cross-validation; finally, the network was tested on 220 pediatric CXR images (P260_220unlabeled dataset).

RESULTS

The distance transform-based bone-edge detection network achieved a mean boundary distance of 1.029. Moreover, the traditional image processing-based bone-suppression model obtained bone images exhibiting a relative Weber contrast of 93.0%. Finally, the fully automated pediatric bone-suppression network achieved a relative mean absolute error of 3.38%, a peak signal-to-noise ratio of 35.5 dB, a structural similarity index measure of 98.1%, and a bone-suppression ratio of 90.1% on P260_40labeled.

CONCLUSIONS

The proposed fully automated pediatric bone-suppression network, together with the proposed distance transform-based bone-edge detection network, can automatically acquire bone and soft-tissue images solely from CXR images for pediatric patients and has the potential to help diagnose pneumonia in children.

Automated Patient-specific Quality Assurance for Automated Segmentation of Organs at Risk in Nasopharyngeal Carcinoma Radiotherapy

INTRODUCTION

Precision radiotherapy relies on accurate segmentation of tumor targets and organs at risk (OARs). Clinicians manually review automatically delineated structures on a case-by-case basis, a time-consuming process dependent on reviewer experience and alertness. This study proposes a general process for automated threshold generation for structural evaluation indicators and patient-specific quality assurance (QA) for automated segmentation of nasopharyngeal carcinoma (NPC).

METHODS

The patient-specific QA process for automated segmentation involves determining the confidence limit and error structure highlight stage. Three expert physicians segmented 17 OARs using computed tomography images of NPC and compared them using the Dice similarity coefficient, the maximum Hausdorff distance, and the mean distance to agreement. For each OAR, the 95% confidence interval was calculated as the confidence limit for each indicator. If two or more evaluation indicators (N2) or one or more evaluation indicators (N1) exceeded the confidence limits, the structure segmentation result was considered abnormal. The quantitative performances of these two methods were compared with those obtained by artificially introducing small/medium and serious errors.

RESULTS

The sensitivity, specificity, balanced accuracy, and F-score values for N2 were 0.944 ± 0.052, 0.827 ± 0.149, 0.886 ± 0.076, and 0.936 ± 0.045, respectively, whereas those for N1 were 0.955 ± 0.045, 0.788 ± 0.189, 0.878 ± 0.096, and 0.948 ± 0.035, respectively. N2 and N1 had small/medium error detection rates of 97.67 ± 0.04% and 98.67 ± 0.04%, respectively, with a serious error detection rate of 100%.

CONCLUSION

The proposed automated patient-specific QA process effectively detected segmentation abnormalities, particularly serious errors. These are crucial for enhancing review efficiency and automated segmentation, and for improving physician confidence in automated segmentation.

Evaluation of Pelvic MRI-to-CT Deformable Registration for Adaptive MR-Guided Particle Therapy

Purpose

We aim to assess the magnetic resonance imaging (MRI)-to-CT deformable image registration (DIR) quality of our treatment planning system in the pelvic region as the first step of an online MRI-guided particle therapy clinical workflow.

Materials and Methods

Using 2 different DIR algorithms, ANAtomically CONstrained Deformation Algorithm (ANACONDA), the DIR algorithm incorporated in RayStation, and Elastix, an open-source registration software, we retrospectively assessed the quality of the deformed CT (dCT) generation in the pelvic region for 5 patients. T1- and T2-weighted daily control MRI acquired prior to treatment delivery were used for the DIR. We compared the contours automatically mapped on the dCT against the manual contours on the MRI (ground truth) by calculating the Dice similarity coefficients and mean distances to the agreement for organs at risk, targets, and outer contour. We assessed the dosimetric impact of the DIR on the clinical treatment plans, comparing the dose-volume histograms and the value of the clinical goals achieved for each dCT. The water equivalent path lengths and dose range 80% (R80%) maps were compared by casting on the beams' eye view.

Results

The T1 sequences performed better for the DIR with ANACONDA compared against the T2. ANACONDA's performance agreed with Elastix. The bladder and rectum led to the worst agreement. For the remaining structures analyzed, Dice similarity coefficients above 0.80 were obtained. Maximum median deviations of 7.1 and 2.1 mm were observed for water equivalent path lengths and R80%, respectively, on the PTV.

Conclusion

This work shows a good agreement on the DIR quality achieved with ANACONDA for the structures in the beams' path. By comparing the R80% generated with ANACONDA and Elastix, we give a first quantification of the uncertainties to be considered in an online MRI-guided particle therapy workflow for pelvic treatment.

Deep Learning-Based Analysis of Aortic Morphology From Three-Dimensional MRI

BACKGROUND

Quantification of aortic morphology plays an important role in the evaluation and follow-up assessment of patients with aortic diseases, but often requires labor-intensive and operator-dependent measurements. Automatic solutions would help enhance their quality and reproducibility.

PURPOSE

To design a deep learning (DL)-based automated approach for aortic landmarks and lumen detection derived from three-dimensional (3D) MRI.

STUDY TYPE

Retrospective.

POPULATION

Three hundred ninety-one individuals (female: 47%, age = 51.9 ± 18.4) from three sites, including healthy subjects and patients (hypertension, aortic dilation, Turner syndrome), randomly divided into training/validation/test datasets (N = 236/77/78). Twenty-five subjects were randomly selected and analyzed by three operators with different levels of expertise.

FIELD STRENGTH/SEQUENCE

1.5-T and 3-T, 3D spoiled gradient-recalled or steady-state free precession sequences.

ASSESSMENT

Reinforcement learning and a two-stage network trained using reference landmarks and segmentation from an existing semi-automatic software were used for aortic landmark detection and segmentation from sinotubular junction to coeliac trunk. Aortic segments were defined using the detected landmarks while the aortic centerline was extracted from the segmentation and morphological indices (length, aortic diameter, and volume) were computed for both the reference and the proposed segmentations.

STATISTICAL TESTS

Segmentation: Dice similarity coefficient (DSC), Hausdorff distance (HD), average symmetrical surface distance (ASSD); landmark detection: Euclidian distance (ED); model robustness: Spearman correlation, Bland-Altman analysis, Kruskal-Wallis test for comparisons between reference and DL-derived aortic indices; inter-observer study: Williams index (WI). A WI 95% confidence interval (CI) lower bound >1 indicates that the method is within the inter-observer variability. A P-value <0.05 was considered statistically significant.

RESULTS

DSC was 0.90 ± 0.05, HD was 12.11 ± 7.79 mm, and ASSD was 1.07 ± 0.63 mm. ED was 5.0 ± 6.1 mm. A good agreement was found between all DL-derived and reference aortic indices (r >0.95, mean bias <7%). Our segmentation and landmark detection performances were within the inter-observer variability except the sinotubular junction landmark (CI = 0.96;1.04).

DATA CONCLUSION

A DL-based aortic segmentation and anatomical landmark detection approach was developed and applied to 3D MRI data for achieve aortic morphology evaluation.

EVIDENCE LEVEL

3 TECHNICAL EFFICACY: Stage 2.

Geometric target margin strategy of proton craniospinal irradiation for pediatric medulloblastoma

In proton craniospinal irradiation (CSI) for skeletally immature pediatric patients, a treatment plan should be developed to ensure that the dose is uniformly delivered to all vertebrae, considering the effects on bone growth balance. The technical (t) clinical target volume (CTV) is conventionally set by manually expanding the CTV from the entire intracranial space and thecal sac, based on the physician's experience. However, there are differences in contouring methods among physicians. Therefore, we aimed to propose a new geometric target margin strategy. Nine pediatric patients with medulloblastoma who underwent proton CSI were enrolled. We measured the following water equivalent lengths for each vertebra in each patient: body surface to the dorsal spinal canal, vertebral limbus, ventral spinal canal and spinous processes. A simulated tCTV (stCTV) was created by assigning geometric margins to the spinal canal using the measurement results such that the vertebral limb and dose distribution coincided with a margin assigned to account for the uncertainty of the proton beam range. The stCTV with a growth factor (correlation between body surface area and age) and tCTV were compared and evaluated. The median values of each index for cervical, thoracic and lumber spine were: the Hausdorff distance, 9.14, 9.84 and 9.77 mm; mean distance-to-agreement, 3.26, 2.65 and 2.64 mm; Dice coefficient, 0.84, 0.81 and 0.82 and Jaccard coefficient, 0.50, 0.60 and 0.62, respectively. The geometric target margin setting method used in this study was useful for creating an stCTV to ensure consistent and uniform planning.

Investigation on ultrasound images for detection of fetal congenital heart defects

Congenital heart defects (CHD) are one of the serious problems that arise during pregnancy. Early CHD detection reduces death rates and morbidity but is hampered by the relatively low detection rates (i.e., 60%) of current screening technology. The detection rate could be increased by supplementing ultrasound imaging with fetal ultrasound image evaluation (FUSI) using deep learning techniques. As a result, the non-invasive foetal ultrasound image has clear potential in the diagnosis of CHD and should be considered in addition to foetal echocardiography. This review paper highlights cutting-edge technologies for detecting CHD using ultrasound images, which involve pre-processing, localization, segmentation, and classification. Existing technique of preprocessing includes spatial domain filter, non-linear mean filter, transform domain filter, and denoising methods based on Convolutional Neural Network (CNN); segmentation includes thresholding-based techniques, region growing-based techniques, edge detection techniques, Artificial Neural Network (ANN) based segmentation methods, non-deep learning approaches and deep learning approaches. The paper also suggests future research directions for improving current methodologies.

CI-UNet: melding convnext and cross-dimensional attention for robust medical image segmentation

Deep learning-based methods have recently shown great promise in medical image segmentation task. However, CNN-based frameworks struggle with inadequate long-range spatial dependency capture, whereas Transformers suffer from computational inefficiency and necessitate substantial volumes of labeled data for effective training. To tackle these issues, this paper introduces CI-UNet, a novel architecture that utilizes ConvNeXt as its encoder, amalgamating the computational efficiency and feature extraction capabilities. Moreover, an advanced attention mechanism is proposed to captures intricate cross-dimensional interactions and global context. Extensive experiments on two segmentation datasets, namely BCSD, and CT2USforKidneySeg, confirm the excellent performance of the proposed CI-UNet as compared to other segmentation methods.

POST-IVUS: A perceptual organisation-aware selective transformer framework for intravascular ultrasound segmentation

Intravascular ultrasound (IVUS) is recommended in guiding coronary intervention. The segmentation of coronary lumen and external elastic membrane (EEM) borders in IVUS images is a key step, but the manual process is time-consuming and error-prone, and suffers from inter-observer variability. In this paper, we propose a novel perceptual organisation-aware selective transformer framework that can achieve accurate and robust segmentation of the vessel walls in IVUS images. In this framework, temporal context-based feature encoders extract efficient motion features of vessels. Then, a perceptual organisation-aware selective transformer module is proposed to extract accurate boundary information, supervised by a dedicated boundary loss. The obtained EEM and lumen segmentation results will be fused in a temporal constraining and fusion module, to determine the most likely correct boundaries with robustness to morphology. Our proposed methods are extensively evaluated in non-selected IVUS sequences, including normal, bifurcated, and calcified vessels with shadow artifacts. The results show that the proposed methods outperform the state-of-the-art, with a Jaccard measure of 0.92 for lumen and 0.94 for EEM on the IVUS 2011 open challenge dataset. This work has been integrated into a software QCU-CMS2 to automatically segment IVUS images in a user-friendly environment.

Optimized Atlas-Based Auto-Segmentation of Bony Structures from Whole-Body Computed Tomography

PURPOSE

To develop and test a method for fully automated segmentation of bony structures from whole-body computed tomography (CT) and evaluate its performance compared with manual segmentation.

METHODS AND MATERIALS

We developed a workflow for automatic whole-body bone segmentation using atlas-based segmentation (ABS) method with a postprocessing module (ABSPP) in MIM MAESTRO software. Fifty-two CT scans comprised the training set to build the atlas library, and 29 CT scans comprised the test set. To validate the workflow, we compared Dice similarity coefficient (DSC), mean distance to agreement, and relative volume errors between ABSPP and ABS with no postprocessing (ABSNPP) with manual segmentation as the reference (gold standard).

RESULTS

The ABSPP method resulted in significantly improved segmentation accuracy (DSC range, 0.85-0.98) compared with the ABSNPP method (DSC range, 0.55-0.87; P < .001). Mean distance to agreement results also indicated high agreement between ABSPP and manual reference delineations (range, 0.11-1.56 mm), which was significantly improved compared with ABSNPP (range, 1.00-2.34 mm) for the majority of tested bony structures. Relative volume errors were also significantly lower for ABSPP compared with ABSNPP for most bony structures.

CONCLUSIONS

We developed a fully automated MIM workflow for bony structure segmentation from whole-body CT, which exhibited high accuracy compared with manual delineation. The integrated postprocessing module significantly improved workflow performance.

A survey on deep learning for skin lesion segmentation

Skin cancer is a major public health problem that could benefit from computer-aided diagnosis to reduce the burden of this common disease. Skin lesion segmentation from images is an important step toward achieving this goal. However, the presence of natural and artificial artifacts (e.g., hair and air bubbles), intrinsic factors (e.g., lesion shape and contrast), and variations in image acquisition conditions make skin lesion segmentation a challenging task. Recently, various researchers have explored the applicability of deep learning models to skin lesion segmentation. In this survey, we cross-examine 177 research papers that deal with deep learning-based segmentation of skin lesions. We analyze these works along several dimensions, including input data (datasets, preprocessing, and synthetic data generation), model design (architecture, modules, and losses), and evaluation aspects (data annotation requirements and segmentation performance). We discuss these dimensions both from the viewpoint of select seminal works, and from a systematic viewpoint, examining how those choices have influenced current trends, and how their limitations should be addressed. To facilitate comparisons, we summarize all examined works in a comprehensive table as well as an interactive table available online³.

Impact of MRI resolution for Linac-based stereotactic radiosurgery

Objective

Magnetic resonance imaging (MRI) is a standard imaging modality in intracranial stereotactic radiosurgery (SRS) for defining target volumes. However, wide disparities in MRI resolution exist, which could directly impact accuracy of target delineation. Here, sequences with various MRI resolution were acquired on phantoms to evaluate the effect on volume definition and dosimetric consequence for cranial SRS.

Materials/Methods

Four T1-weighted MR sequences with increasing 3D resolution were compared, including two Spin Echo (SE) 2D acquisitions with 5mm and 3mm slice thickness (SE5mm, SE3mm) and two gradient echo 3D acquisitions (TFE, BRAVO). The voxel sizes were 0.4×0.4×5.0, 0.5×0.5×3.0, 0.9×0.9×1.25, and 0.4×0.4×0.5 mm³, respectively. Four phantoms with simulated lesions of different shape and volume (range, 0.53-25.0 cm³) were imaged, resulting in 16 total sets of MRIs. Four radiation oncologists provided contours on individual MR image set. All observer contours were compared with ground truth, defined on CT image according to the absolute dimensions of the target structure, using Dice similarity coefficient (DSC), Hausdorff distance (HD), mean distance-to-agreement (MDA), and the ratio between reconstructed and true volume (Ratio_vol ). For dosimetric consequence, SRS plans targeting observer volumes were created. The true Paddick conformity index ( C I p a d d i c k t r u e ), calculated with true target volume, was correlated with quality of observer volume.

Results

All measures of observer contours improved as increasingly higher MRI resolution was provided from SE5mm to BRAVO. The improvement in DSC, HD and MDA was statistically significant (p<0.01). Dosimetrically, C I p a d d i c k t r u e   strongly correlated with DSC of the planning observer volume (Pearson's r=0.94, p<0.00001).

Conclusions

Significant improvement in target definition and reduced inter-observer variation was observed as the MRI resolution improved, which also improved the quality of SRS plans. Results imply that high resolution 3D MR sequences should be used to minimize potential errors in target definition, and multi-slice 2D sequences should be avoided.

Comparison of two-dimensional and three-dimensional U-Net architectures for segmentation of adipose tissue in cardiac magnetic resonance images

The process of identifying cardiac adipose tissue (CAT) from volumetric magnetic resonance imaging of the heart is tedious, time-consuming, and often dependent on observer interpretation. Many 2-dimensional (2D) convolutional neural networks (CNNs) have been implemented to automate the cardiac segmentation process, but none have attempted to identify CAT. Furthermore, the results from automatic segmentation of other cardiac structures leave room for improvement. This study investigated the viability of a 3-dimensional (3D) CNN in comparison to a similar 2D CNN. Both models used a U-Net architecture to simultaneously classify CAT, left myocardium, left ventricle, and right myocardium. The multi-phase model trained with multiple observers' segmentations reached a whole-volume Dice similarity coefficient (DSC) of 0.925 across all classes and 0.640 for CAT specifically; the corresponding 2D model's DSC across all classes was 0.902 and 0.590 for CAT specifically. This 3D model also achieved a higher level of CAT-specific DSC agreement with a group of observers with a Williams Index score of 0.973 in comparison to the 2D model's score of 0.822.

Single projection driven real-time multi-contrast (SPIDERM) MR imaging using pre-learned spatial subspace and linear transformation

Objective. To develop and test the feasibility of a novel Single ProjectIon DrivEn Real-time Multi-contrast (SPIDERM) MR imaging technique that can generate real-time 3D images on-the-fly with flexible contrast weightings and a low latency. Approach. In SPIDERM, a ‘prep’ scan is first performed, with sparse k-space sampling periodically interleaved with the central k-space line (navigator data), to learn a subject-specific model, incorporating a spatial subspace and a linear transformation between navigator data and subspace coordinates. A ‘live’ scan is then performed by repeatedly acquiring the central k-space line only to dynamically determine subspace coordinates. With the ‘prep’-learned subspace and ‘live’ coordinates, real-time 3D images are generated on-the-fly with computationally efficient matrix multiplication. When implemented based on a multi-contrast pulse sequence, SPIDERM further allows for data-driven image contrast regeneration to convert real-time contrast-varying images into contrast-frozen images at user’s discretion while maintaining motion states. Both digital phantom and in-vivo experiments were performed to evaluate the technical feasibility of SPIDERM. Main results. The elapsed time from the input of the central k-space line to the generation of real-time contrast-frozen 3D images was approximately 45 ms, permitting a latency of 55 ms or less. Motion displacement measured from SPIDERM and reference images showed excellent correlation ( R 2 ≥ 0.983 ). Geometric variation from the ground truth in the digital phantom was acceptable as demonstrated by pancreas contour analysis (Dice ≥ 0.84, mean surface distance ≤ 0.95 mm). Quantitative image quality metrics showed good consistency between reference images and contrast-varying SPIDREM images in in-vivo studies (mean NMRSE = 0.141 , PSNR = 3 0.12 , SSIM = 0.88 ). Significance. SPIDERM is capable of generating real-time multi-contrast 3D images with a low latency. An imaging framework based on SPIDERM has the potential to serve as a standalone package for MR-guided radiation therapy by offering adaptive simulation through a ‘prep’ scan and real-time image guidance through a ‘live’ scan.

Pulmonary Lesion Classification Framework Using the Weighted Ensemble Classification with Random Forest and CNN Models for EBUS Images

Lung cancer is a deadly disease with a high mortality rate. Endobronchial ultrasonography (EBUS) is one of the methods for detecting pulmonary lesions. Computer-aided diagnosis of pulmonary lesions from images can help radiologists to classify lesions; however, most of the existing methods need a large volume of data to give good results. Thus, this paper proposes a novel pulmonary lesion classification framework for EBUS images that works well with small datasets. The proposed framework integrates the statistical results from three classification models using the weighted ensemble classification. The three classification models include the radiomics feature and patient data-based model, the single-image-based model, and the multi-patch-based model. The radiomics features are combined with the patient data to be used as input data for the random forest, whereas the EBUS images are used as input data to the other two CNN models. The performance of the proposed framework was evaluated on a set of 200 EBUS images consisting of 124 malignant lesions and 76 benign lesions. The experimental results show that the accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and area under the curve are 95.00%, 100%, 86.67%, 92.59%, 100%, and 93.33%, respectively. This framework can significantly improve the pulmonary lesion classification.

MIRSIG position paper: the use of image registration and fusion algorithms in radiotherapy

The report of the American Association of Physicists in Medicine (AAPM) Task Group No. 132 published in 2017 reviewed rigid image registration and deformable image registration (DIR) approaches and solutions to provide recommendations for quality assurance and quality control of clinical image registration and fusion techniques in radiotherapy. However, that report did not include the use of DIR for advanced applications such as dose warping or warping of other matrices of interest. Considering that DIR warping tools are now readily available, discussions were hosted by the Medical Image Registration Special Interest Group (MIRSIG) of the Australasian College of Physical Scientists & Engineers in Medicine in 2018 to form a consensus on best practice guidelines. This position statement authored by MIRSIG endorses the recommendations of the report of AAPM task group 132 and expands on the best practice advice from the 'Deforming to Best Practice' MIRSIG publication to provide guidelines on the use of DIR for advanced applications.

Clinical target segmentation using a novel deep neural network: double attention Res-U-Net

We introduced Double Attention Res-U-Net architecture to address medical image segmentation problem in different medical imaging system. Accurate medical image segmentation suffers from some challenges including, difficulty of different interest object modeling, presence of noise, and signal dropout throughout the measurement. The base line image segmentation approaches are not sufficient for complex target segmentation throughout the various medical image types. To overcome the issues, a novel U-Net-based model proposed that consists of two consecutive networks with five and four encoding and decoding levels respectively. In each of networks, there are four residual blocks between the encoder-decoder path and skip connections that help the networks to tackle the vanishing gradient problem, followed by the multi-scale attention gates to generate richer contextual information. To evaluate our architecture, we investigated three distinct data-sets, (i.e., CVC-ClinicDB dataset, Multi-site MRI dataset, and a collected ultrasound dataset). The proposed algorithm achieved Dice and Jaccard coefficients of 95.79%, 91.62%, respectively for CRL, and 93.84% and 89.08% for fetal foot segmentation. Moreover, the proposed model outperformed the state-of-the-art U-Net based model on the external CVC-ClinicDB, and multi-site MRI datasets with Dice and Jaccard coefficients of 83%, 75.31% for CVC-ClinicDB, and 92.07% and 87.14% for multi-site MRI dataset, respectively.

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

•

Convolutional neural networks (CNNs) are used for auto-segmentation in radiotherapy.

•

However, CNNs rely on large, high-quality datasets: a scarcity in radiotherapy.

•

We develop a CNN model, trained with limited data, for accurate segmentation.

•

Multiple experiments were performed to optimise key features of our custom model.

•

Our model is competitive with state-of-the-art methods on a public dataset.

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.

Interobserver variability in target volume delineation for CT/MRI simulation and MRI-guided adaptive radiotherapy in rectal cancer

OBJECTIVES

Quantify target volume delineation uncertainty for CT/MRI simulation and MRI-guided adaptive radiotherapy in rectal cancer. Define optimal imaging sequences for target delineation.

METHODS

Six experienced radiation oncologists delineated clinical target volumes (CTVs) on CT and 2D and 3D-MRI in three patients with rectal cancer, using consensus contouring guidelines. Tumour GTV (GTVp) was also contoured on MRI acquired week 0 and 3 of radiotherapy. A STAPLE contour was created and volume and interobserver variability metrics were analysed.

RESULTS

There were statistically significant differences in volume between observers for CT and 2D-MRI-defined CTVs (p < 0.05). There was no significant difference between observers on 3D-MRI. Significant differences in volume were seen between observers for both 2D and 3D-MRI-defined GTVp at weeks 0 and 3 (p < 0.05). Good interobserver agreement (IOA) was seen for CTVs delineated on all imaging modalities with best IOA on 3D-MRI; median Conformity index (CI) 0.74 for CT, 0.75 for 2D-MRI and 0.77 for 3D-MRI. IOA of MRI-defined GTVp week 0 was better compared to CT; CI 0.58 for CT, 0.62 for 2D-MRI and 0.7 for 3D-MRI. MRI-defined GTVp IOA week three was worse compared to week 0.

CONCLUSION

Delineation on MRI results in smaller volumes and better IOA week 0 compared to CT. 3D-MRI provides the best IOA in CTV and GTVp. MRI-defined GTVp on images acquired week 3 showed worse IOA compared to week 0. This highlights the need for consensus guidelines in GTVp delineation on MRI during treatment course in the context of dose escalation MRI-guided rectal boost studies.

ADVANCES IN KNOWLEDGE

Optimal MRI sequences for CT/MRI simulation and MRI-guided adaptive radiotherapy in rectal cancer have been defined.

Assessment of CT to CBCT contour mapping for radiomic feature analysis in prostate cancer

This study provides a quantitative assessment of the accuracy of a commercially available deformable image registration (DIR) algorithm to automatically generate prostate contours and additionally investigates the robustness of radiomic features to differing contours. Twenty-eight prostate cancer patients enrolled on an institutional review board (IRB) approved protocol were selected. Planning CTs (pCTs) were deformably registered to daily cone-beam CTs (CBCTs) to generate prostate contours (auto contours). The prostate contours were also manually drawn by a physician. Quantitative assessment of deformed versus manually drawn prostate contours on daily CBCT images was performed using Dice similarity coefficient (DSC), mean distance-to-agreement (MDA), difference in center-of-mass position (ΔCM) and difference in volume (ΔVol). Radiomic features from 6 classes were extracted from each contour. Lin’s concordance correlation coefficient (CCC) and mean absolute percent difference in radiomic feature-derived data (mean |%Δ|RF) between auto and manual contours were calculated. The mean (± SD) DSC, MDA, ΔCM and ΔVol between the auto and manual prostate contours were 0.90 ± 0.04, 1.81 ± 0.47 mm, 2.17 ± 1.26 mm and 5.1 ± 4.1% respectively. Of the 1,010 fractions under consideration, 94.8% of DIRs were within TG-132 recommended tolerance. 30 radiomic features had a CCC > 0.90 and 21 had a mean |%∆|RF < 5%. Auto-propagation of prostate contours resulted in nearly 95% of DIRs within tolerance recommendations of TG-132, leading to the majority of features being regarded as acceptably robust. The use of auto contours for radiomic feature analysis is promising but must be done with caution.

A CNN-based method to reconstruct 3-D spine surfaces from US images in vivo

Three-dimensional (3-D) reconstruction of the spine surface is of strong clinical relevance for the diagnosis and prognosis of spine disorders and intra-operative image guidance. In this paper, we report a new technique to reconstruct lumbar spine surfaces in 3-D from non-invasive ultrasound (US) images acquired in free-hand mode. US images randomly sampled from in vivo scans of 9 rabbits were used to train a U-net convolutional neural network (CNN). More specifically, a late fusion (LF)-based U-net trained jointly on B-mode and shadow-enhanced B-mode images was generated by fusing two individual U-nets and expanding the set of trainable parameters to around twice the capacity of a basic U-net. This U-net was then applied to predict spine surface labels in in vivo images obtained from another rabbit, which were then used for 3-D spine surface reconstruction. The underlying pose of the transducer during the scan was estimated by registering stacks of US images to a geometrical model derived from corresponding CT data and used to align detected surface points. Final performance of the reconstruction method was assessed by computing the mean absolute error (MAE) between pairs of spine surface points detected from US and CT and by counting the total number of surface points detected from US. Comparison was made between the LF-based U-net and a previously developed phase symmetry (PS)-based method. Using the LF-based U-net, the averaged number of US surface points across the lumbar region increased by 21.61% and MAE reduced by 26.28% relative to the PS-based method. The overall MAE (in mm) was 0.24±0.29. Based on these results, we conclude that: 1) the proposed U-net can detect the spine posterior arch with low MAE and large number of US surface points and 2) the newly proposed reconstruction framework may complement and, under certain circumstances, be used without the aid of an external tracking system in intra-operative spine applications.

Evaluation of performance of pelvic CT-MR deformable image registration using two software programs

We assessed the accuracy of deformable image registration (DIR) accuracy between CT and MR images using an open-source software (Elastix, from Utrecht Medical Center) and a commercial software (Velocity AI Ver. 3.2.0 from Varian Medical Systems, Palo Alto, CA, USA) software. Five male patients' pelvic regions were studied using publicly available CT, T1-weighted (T1w) MR, and T2-weighted (T2w) MR images. In the cost function of the Elastix, we used six DIR parameter settings with different regularization weights (Elastix0, Elastix0.01, Elastix0.1, Elastix1, Elastix10, and Elastix100). We used MR Corrected Deformable algorithm for Velocity AI. The Dice similarity coefficient (DSC) and mean distance to agreement (MDA) for the prostate, bladder, rectum and left and right femoral heads were used to evaluate DIR accuracy. Except for the bladder, most algorithms produced good DSC and MDA results for all organs. In our study, the mean DSCs for the bladder ranged from 0.75 to 0.88 (CT-T1w) and from 0.72 to 0.76 (CT-T2w). Similarly, the mean MDA ranges were 2.4 to 4.9 mm (CT-T1w), 4.6 to 5.3 mm (CT-T2w). For the Elastix, CT-T1w was outperformed CT-T2w for both DSCs and MDAs at Elastix0, Elastix0.01, and Elastix0.1. In the case of Velocity AI, no significant differences in DSC and MDA of all organs were observed. This implied that the DIR accuracy of CT and MR images might differ depending on the sequence used.

Deformable image registration to assist clinical decision for radiotherapy treatment adaptation for head and neck cancer patients

Head and neck (H&N) cancer patients often present anatomical and geometrical changes in tumors and organs at risk (OARs) during radiotherapy treatment. These changes may result in the need to adapt the existing treatment planning, using an expert's subjective opinion, for offline adaptive radiotherapy and a new treatment planning before each treatment, for online adaptive radiotherapy. In the present study, a fast methodology is proposed to assist in planning adaptation clinical decision using tumor and parotid glands percentage volume changes during treatment. The proposed approach was applied to 40 Η&Ν cases, with one planning Computed Tomography (pCT) image and CBCT scans for 6 weeks of treatment per case. Deformable registration was used for each patient's pCT image alignment to its weekly CBCT. The calculated transformations were used to align each patient's anatomical structures to the weekly anatomy. Clinical target volume (CTV) and parotid gland volume percentage changes were calculated in each case. The accuracy of the achieved image alignment was validated qualitatively and quantitatively. Furthermore, statistical analysis was performed to test if there is a statistically significant correlation between CTV and parotid glands volume percentage changes. Average MDA for CTV and parotid glands between corresponding structures defined by an expert in CBCTs and automatically calculated through registration was 1.4 ± 0.1 mm and 1.5 ± 0.1 mm, respectively. The mean registration time of the first CBCT image registration for 40 cases was lower than 3.4 min. Five patients show more than 20% tumor volume change. Six patients show more than 30% parotid glands volume change. Ten out of 40 patients proposed for planning adaptation. All the statistical tests performed showed no correlation between CTV/parotid glands percentage volume changes. The aim to assist in clinical decision making on a fast and automatic way was achieved using the proposed methodology, thereby reducing workload in clinical practice.

Rigid and Deformable Image Registration for Radiation Therapy: A Self-Study Evaluation Guide for NRG Oncology Clinical Trial Participation

PURPOSE

The registration of multiple imaging studies to radiation therapy computed tomography simulation, including magnetic resonance imaging, positron emission tomography-computed tomography, etc. is a widely used strategy in radiation oncology treatment planning, and these registrations have valuable roles in image guidance, dose composition/accumulation, and treatment delivery adaptation. The NRG Oncology Medical Physics subcommittee formed a working group to investigate feasible workflows for a self-study credentialing process of image registration commissioning.

METHODS AND MATERIALS

The American Association of Physicists in Medicine (AAPM) Task Group 132 (TG132) report on the use of image registration and fusion algorithms in radiation therapy provides basic guidelines for quality assurance and quality control of the image registration algorithms and the overall clinical process. The report recommends a series of tests and the corresponding metrics that should be evaluated and reported during commissioning and routine quality assurance, as well as a set of recommendations for vendors. The NRG Oncology medical physics subcommittee working group found incompatibility of some digital phantoms with commercial systems. Thus, there is still a need to provide further recommendations in terms of compatible digital phantoms, clinical feasible workflow, and achievable thresholds, especially for future clinical trials involving deformable image registration algorithms. Nine institutions participated and evaluated 4 commonly used commercial imaging registration software and various versions in the field of radiation oncology.

RESULTS AND CONCLUSIONS

The NRG Oncology Working Group on image registration commissioning herein provides recommendations on the use of digital phantom/data sets and analytical software access for institutions and clinics to perform their own self-study evaluation of commercial imaging systems that might be employed for coregistration in radiation therapy treatment planning and image guidance procedures. Evaluation metrics and their corresponding values were given as guidelines to establish practical tolerances. Vendor compliance for image registration commissioning was evaluated, and recommendations were given for future development.

Advanced deep learning methodology for accurate, real-time segmentation of high-resolution intravascular ultrasound images

AIMS

The aim of this study is to develop and validate a deep learning (DL) methodology capable of automated and accurate segmentation of intravascular ultrasound (IVUS) image sequences in real-time.

METHODS AND RESULTS

IVUS segmentation was performed by two experts who manually annotated the external elastic membrane (EEM) and lumen borders in the end-diastolic frames of 197 IVUS sequences portraying the native coronary arteries of 65 patients. The IVUS sequences of 177 randomly-selected vessels were used to train and optimise a novel DL model for the segmentation of IVUS images. Validation of the developed methodology was performed in 20 vessels using the estimations of two expert analysts as the reference standard. The mean difference for the EEM, lumen and plaque area between the DL-methodology and the analysts was ≤0.23mm² (standard deviation ≤0.85mm²), while the Hausdorff and mean distance differences for the EEM and lumen borders was ≤0.19 mm (standard deviation≤0.17 mm). The agreement between DL and experts was similar to experts' agreement (Williams Index ranges: 0.754-1.061) with similar results in frames portraying calcific plaques or side branches.

CONCLUSIONS

The developed DL-methodology appears accurate and capable of segmenting high-resolution real-world IVUS datasets. These features are expected to facilitate its broad adoption and enhance the applications of IVUS in clinical practice and research.

Automatic segmentation and tracking of biological prosthetic heart valves

Purpose: Prosthetic heart valve designs must be rigorously tested using cardiovascular equipment. The valve orifice area over time constitutes a key quality metric which is typically assessed manually, thus a tedious and error-prone task. From a computer vision viewpoint, a major unsolved issue lies in the orifice being partly occluded by the leaflets' inner side or inaccurately depicted due to its transparency. Here, we address this issue, which allows us to focus on the accurate and automatic computation of valve orifice areas. Approach: We propose a segmentation approach based on the detection of the leaflets' free edges. Using video frames recorded with a high-speed digital camera during in vitro simulations, an initial estimation of the orifice area is first obtained via active contouring and thresholding and then refined to capture the leaflet free edges via a curve transformation mechanism. Results: Experiments on video data from pulsatile flow testing demonstrate the effectiveness of our approach: a root-mean-square error (RMSE) on the temporal extracted orifice areas between 0.8% and 1.2%, an average Jaccard similarity coefficient between 0.933 and 0.956, and an average Hausdorff distance between 7.2 and 11.9 pixels. Conclusions: Our approach significantly outperformed a state-of-the-art algorithm in terms of evaluation metrics related to valve design (RMSE) and computer vision (accuracy of the orifice shape). It can also cope with lower quality videos and is better at processing frames showing an almost closed valve, a crucial quality for assessing valve design malfunctions related to their improper closing.

Cranial organs at risk delineation: heterogenous practices in radiotherapy planning

BACKGROUND

Segmentation is a crucial step in treatment planning that directly impacts dose distribution and optimization. The aim of this study was to evaluate the inter-individual variability of common cranial organs at risk (OAR) delineation in neurooncology practice.

METHODS

Anonymized simulation contrast-enhanced CT and MR scans of one patient with a solitary brain metastasis was used for delineation and analysis. Expert professionals from 16 radiotherapy centers involved in brain structures delineation were asked to segment 9 OAR on their own treatment planning system. As reference, two experts in neurooncology, produced a unique consensual contour set according to guidelines. Overlap ratio, Kappa index (KI), volumetric ratio, Commonly Contoured Volume, Supplementary Contoured Volume were evaluated using Artiview™ v 2.8.2-according to occupation, seniority and level of expertise of all participants.

RESULTS

For the most frequently delineated and largest OAR, the mean KI are often good (0.8 for the parotid and the brainstem); however, for the smaller OAR, KI degrade (0.3 for the optic chiasm, 0.5% for the cochlea), with a significant discrimination (p < 0.01). The radiation oncologists, members of Association des Neuro-Oncologue d'Expression Française society performed better in all indicators compared to non-members (p < 0.01). Our exercise was effective in separating the different participating centers with 3 of the reported indicators (p < 0.01).

CONCLUSION

Our study illustrates the heterogeneity in normal structures contouring between professionals. We emphasize the need for cerebral OAR delineation harmonization-that is a major determinant of therapeutic ratio and clinical trials evaluation.

Automated 3D geometry segmentation of the healthy and diseased carotid artery in free-hand, probe tracked ultrasound images

PURPOSE

Rupture of an arterosclerotic plaque in the carotid artery is a major cause of stroke. Biomechanical analysis of plaques is under development aiming to aid the clinician in the assessment of plaque vulnerability. Patient-specific three-dimensional (3D) geometry assessment of the carotid artery, including the bifurcation, is required as input for these biomechanical models. This requires a high-resolution, 3D, noninvasive imaging modality such as ultrasound (US). In this study, a high-resolution two-dimensional (2D) linear array in combination with a magnetic probe tracking device and automatic segmentation method was used to assess the geometry of the carotid artery. The advantages of using this system over a 3D ultrasound probe are its higher resolution (spatial and temporal) and its larger field of view.

METHODS

A slow sweep (v = ± 5 mm/s) was made over the subject's neck so that the full geometry of the bifurcated geometry of the carotid artery is captured. An automated segmentation pipeline was developed. First, the Star-Kalman method was used to approximate the center and size of the vessels for every frame. Images were filtered with a Gaussian high-pass filter before conversion into the 2D monogenic signals, and multiscale asymmetry features were extracted from these data, enhancing low lateral wall-lumen contrast. These images, in combination with the initial ellipse contours, were used for an active deformable contour model to segment the vessel lumen. To segment the lumen-plaque boundary, Otsu's automatic thresholding method was used. Distension of the wall due to the change in blood pressure was removed using a filter approach. Finally, the contours were converted into a 3D hexahedral mesh for a patient-specific solid mechanics model of the complete arterial wall.

RESULTS

The method was tested on 19 healthy volunteers and on 3 patients. The results were compared to manual segmentation performed by three experienced observers. Results showed an average Hausdorff distance of 0.86 mm and an average similarity index of 0.91 for the common carotid artery (CCA) and 0.88 for the internal and external carotid artery. For the total algorithm, the success rate was 89%, in 4 out of 38 datasets the ICA and ECA were not sufficient visible in the US images. Accurate 3D hexahedral meshes were successfully generated from the segmented images .

CONCLUSIONS

With this method, a subject-specific biomechanical model can be constructed directly from a hand-held 2D US measurement, within 10 min, with a minimal user input. The performance of the proposed segmentation algorithm is comparable to or better than algorithms previously described in literature. Moreover, the algorithm is able to segment the CCA, ICA, and ECA including the carotid bifurcation in transverse B-mode images in both healthy and diseased arteries.

Evaluation of the accuracy of deformable image registration on MRI with a physical phantom

Background and purpose

Magnetic resonance imaging (MRI) has gained popularity in radiation therapy simulation because it provides superior soft tissue contrast, which facilitates more accurate target delineation compared with computed tomography (CT) and does not expose the patient to ionizing radiation. However, image registration errors in commercial software have not been widely reported. Here we evaluated the accuracy of deformable image registration (DIR) by using a physical phantom for MRI.

Methods and materials

We used the “Wuphantom” for end‐to‐end testing of DIR accuracy for MRI. This acrylic phantom is filled with water and includes several fillable inserts to simulate various tissue shapes and properties. Deformations and changes in anatomic locations are simulated by changing the rotations of the phantom and inserts. We used Varian Velocity DIR software (v4.0) and CT (head and neck protocol) and MR (T1‐ and T2‐weighted head protocol) images to test DIR accuracy between image modalities (MRI vs CT) and within the same image modality (MRI vs MRI) in 11 rotation deformation scenarios. Large inserts filled with Mobil DTE oil were used to simulate fatty tissue, and small inserts filled with agarose gel were used to simulate tissues slightly denser than water (e.g., prostate). Contours of all inserts were generated before DIR to provide a baseline for contour size and shape. DIR was done with the MR Correctable Deformable DIR method, and all deformed contours were compared with the original contours. The Dice similarity coefficient (DSC) and mean distance to agreement (MDA) were used to quantitatively validate DIR accuracy. We also used large and small regions of interest (ROIs) during between‐modality DIR tests to simulate validation of DIR accuracy for organs at risk (OARs) and propagation of individual clinical target volume (CTV) contours.

Results

No significant differences in DIR accuracy were found for T1:T1 and T2:T2 comparisons (P > 0.05). DIR was less accurate for between‐modality comparisons than for same‐modality comparisons, and was less accurate for T1 vs CT than for T2 vs CT (P < 0.001). For between‐modality comparisons, use of a small ROI improved DIR accuracy for both T1 and T2 images.

Conclusion

The simple design of the Wuphantom allows seamless testing of DIR; here we validated the accuracy of MRI DIR in end‐to‐end testing. T2 images had superior DIR accuracy compared with T1 images. Use of small ROIs improves DIR accuracy for target contour propagation.

Quantifying the accuracy of deformable image registration for cone-beam computed tomography with a physical phantom

PURPOSE

Kilo-voltage cone-beam computed tomography (CBCT) is widely used for patient alignment, contour propagation, and adaptive treatment planning in radiation therapy. In this study, we evaluated the accuracy of deformable image registration (DIR) for CBCT under various imaging protocols with different noise and patient dose levels.

METHODS

A physical phantom previously developed to facilitate end-to-end testing of the DIR accuracy was used with Varian Velocity v4.0 software to evaluate the performance of image registration from CT to CT, CBCT to CT, and CBCT to CBCT. The phantom is acrylic and includes several inserts that simulate different tissue shapes and properties. Deformations and anatomic changes were simulated by changing the rotations of both the phantom and the inserts. CT images (from a head and neck protocol) and CBCT images (from pelvis, head and "Image Gently" protocols) were obtained with different image noise and dose levels. Large inserts were filled with Mobil DTE oil to simulate soft tissue, and small inserts were filled with bone materials. All inserts were contoured before the DIR process to provide a ground truth contour size and shape for comparison. After the DIR process, all deformed contours were compared with the originals using Dice similarity coefficient (DSC) and mean distance to agreement (MDA). Both large and small volume of interests (VOIs) for DIR volume selection were tested by simulating a DIR process that included whole patient image volume and clinical target volumes (CTV) only (for CTVs propagation).

RESULTS

For cross-modality DIR registration (CT to CBCT), the DSC were >0.8 and the MDA were <3 mm for CBCT pelvis, and CBCT head protocols. For CBCT to CBCT and CT to CT, the DIR accuracy was improved relative to the cross-modality tests. For smaller VOIs, the DSC were >0.8 and MDA <2 mm for all modalities.

CONCLUSIONS

The accuracy of DIR depends on the quality of the CBCT image at different dose and noise levels.

Lumen-intima and media-adventitia segmentation in IVUS images using supervised classifications of arterial layers and morphological structures

BACKGROUND

Intravascular ultrasound (IVUS) provides axial grey-scale images of blood vessels. The large number of images require automatic analysis, specifically to identify the lumen and outer vessel wall. However, the high amount of noise, the presence of artifacts and anatomical structures, such as bifurcations, calcifications and fibrotic plaques, usually hinder the proper automatic segmentation of the vessel wall.

METHODS

Lumen, media, adventitia and surrounding tissues are automatically detected using Support Vector Machines (SVMs). The classification performance of the SVMs vary according to the kind of structure present within each region of the image. Random Forest (RF) is used to detect different morphological structures and to modify the initial layer classification depending on the detected structure. The resulting classification maps are fed into a segmentation method based on deformable contours to detect lumen-intima (LI) and media-adventitia (MA) interfaces.

RESULTS

The modifications in the layer classifications according to the presence of structures proved to be effective improving LI and MA segmentations. The proposed method reaches a Jaccard Measure (JM) of 0.88 ± 0.08 for LI segmentation, compared with 0.88 ± 0.05 of a semiautomatic method. When looking at MA, our method reaches a JM of 0.84 ± 0.09, and outperforms previous automatic methods in terms of HD, with 0.51mm ± 0.30.

CONCLUSIONS

A simple modification to the arterial layer classification produces results that match and improve state-of-the-art fully-automatic segmentation methods for LI and MA in 20MHz IVUS images. For LI segmentation, the proposed automatic method performs accurately as semi-automatic methods. For MA segmentation, our method matched the quality of state-of-the-art automatic methods described in the literature. Furthermore, our implementation is modular and open-source, allowing for future extensions and improvements.

Harmonization of practices between radiotherapy centres in the Nord and Pas-de-Calais regions (France): A three-year evaluation

PURPOSE

The delineation of volumes of interest can be a source of significant interobserver variability. The purpose of this study was to improve the homogeneity of delineation between oncologist-radiotherapists in the territorial departments of Nord and Pas-de-Calais (France) through discussions of clinical cases and the adoption of common published reference documents.

MATERIALS AND METHODS

All eleven radiotherapy centres in the Nord and Pas-de-Calais departments of France participated. The localizations assessed to date included prostate, head and neck, breast and brain cancers. For each localization, the junior or senior physician(s) in charge of pathology delineated the volumes of interest according to their usual practices. Validated indices, including the Dice similarity coefficient, were used to quantify the delineation differences. The anonymized results were presented at two to three annual meetings. A second delineation of the clinical cases was then carried out to quantify homogenization. An evaluation of dosimetry practices was also conducted for prostate cancer. Wilcoxon assay matched data were used.

RESULTS

Our work showed either satisfactory delineation concordance after the initial assessment or improved delineation concordance. For prostate cancer, the Dice similarity coefficient values were greater than 0.6 initially in two of the three clinical cases. For head and neck cancers, a statistically significant improvement was observed for only one of the clinical target volumes. More than half of the Dice similarity coefficient values were greater than 0.6 in the first comparison. The study of clinical cases of breast cancer allowed a homogenization of the delineation of five of the six lymph node clinical target volumes. The dosimetry study of prostate cancer allowed for a homogenization of practices.

CONCLUSION

This work makes it possible to harmonize the delineation practices around validated standards. An extension to the entire Hauts-de-France region is planned.

Impact of 90Y PET gradient-based tumor segmentation on voxel-level dosimetry in liver radioembolization

BACKGROUND

The purpose was to validate ⁹⁰Y PET gradient-based tumor segmentation in phantoms and to evaluate the impact of the segmentation method on reported tumor absorbed dose (AD) and biological effective dose (BED) in ⁹⁰Y microsphere radioembolization (RE) patients. A semi-automated gradient-based method was applied to phantoms and patient tumors on the ⁹⁰Y PET with the initial bounding volume for gradient detection determined from a registered diagnostic CT or MR; this PET-based segmentation (PS) was compared with radiologist-defined morphologic segmentation (MS) on CT or MRI. AD and BED volume histogram metrics (D90, D70, mean) were calculated using both segmentations and concordance/correlations were investigated. Spatial concordance was assessed using Dice similarity coefficient (DSC) and mean distance to agreement (MDA). PS was repeated to assess intra-observer variability.

RESULTS

In phantoms, PS demonstrated high accuracy in lesion volumes (within 15%), AD metrics (within 11%), high spatial concordance relative to morphologic segmentation (DSC > 0.86 and MDA < 1.5 mm), and low intra-observer variability (DSC > 0.99, MDA < 0.2 mm, AD/BED metrics within 2%). For patients (58 lesions), spatial concordance between PS and MS was degraded compared to in-phantom (average DSC = 0.54, average MDA = 4.8 mm); the average mean tumor AD was 226 ± 153 and 197 ± 138 Gy, respectively for PS and MS. For patient AD metrics, the best Pearson correlation (r) and concordance correlation coefficient (ccc) between segmentation methods was found for mean AD (r = 0.94, ccc = 0.92), but worsened as the metric approached the minimum dose (for D90, r = 0.77, ccc = 0.69); BED metrics exhibited a similar trend. Patient PS showed low intra-observer variability (average DSC = 0.81, average MDA = 2.2 mm, average AD/BED metrics within 3.0%).

CONCLUSIONS

⁹⁰Y PET gradient-based segmentation led to accurate/robust results in phantoms, and showed high concordance with MS for reporting mean tumor AD/BED in patients. However, tumor coverage metrics such as D90 exhibited worse concordance between segmentation methods, highlighting the need to standardize segmentation methods when reporting AD/BED metrics from post-therapy ⁹⁰Y PET. Estimated differences in reported AD/BED metrics due to segmentation method will be useful for interpreting RE dosimetry results in the literature including tumor response data.

Evaluation of Image Registration Accuracy for Tumor and Organs at Risk in the Thorax for Compliance With TG 132 Recommendations

Purpose

To evaluate accuracy for 2 deformable image registration methods (in-house B-spline and MIM freeform) using image pairs exhibiting changes in patient orientation and lung volume and to assess the appropriateness of registration accuracy tolerances proposed by the American Association of Physicists in Medicine Task Group 132 under such challenging conditions via assessment by expert observers.

Methods and Materials

Four-dimensional computed tomography scans for 12 patients with lung cancer were acquired with patients in prone and supine positions. Tumor and organs at risk were delineated by a physician on all data sets: supine inhale (SI), supine exhale, prone inhale, and prone exhale. The SI image was registered to the other images using both registration methods. All SI contours were propagated using the resulting transformations and compared with physician delineations using Dice similarity coefficient, mean distance to agreement, and Hausdorff distance. Additionally, propagated contours were anonymized along with ground-truth contours and rated for quality by physician-observers.

Results

Averaged across all patients, the accuracy metrics investigated remained within tolerances recommended by Task Group 132 (Dice similarity coefficient >0.8, mean distance to agreement <3 mm). MIM performed better with both complex (vertebrae) and low-contrast (esophagus) structures, whereas the in-house method performed better with lungs (whole and individual lobes). Accuracy metrics worsened but remained within tolerances when propagating from supine to prone; however, the Jacobian determinant contained regions with negative values, indicating localized nonphysiologic deformations. For MIM and in-house registrations, 50% and 43.8%, respectively, of propagated contours were rated acceptable as is and 8.2% and 11.0% as clinically unacceptable.

Conclusions

The deformable image registration methods performed reliably and met recommended tolerances despite anatomically challenging cases exceeding typical interfraction variability. However, additional quality assurance measures are necessary for complex applications (eg, dose propagation). Human review rather than unsupervised implementation should always be part of the clinical registration workflow.

LEFMIS: locally-oriented evaluation framework for medical image segmentation algorithms

This article proposes a novel framework for the locally-oriented evaluation of segmentation algorithms (LEFMIS). The presented approach is robust and takes into account local inter/intra-observer variability and the anisotropy of medical images. What is more, the framework makes it possible to distinguish types of error locally. These features are crucial in the context of cancer image data. The proposed framework is based on use of the signed anisotropic Euclidean distance transform and the distance projection. It can be used easily in many different applications with or without additional expert outlines (both inter- and intra-observer variability). The performance of the proposed framework is depicted using both artificial and kidney cancer CT data with experts' manual outlines. In the article, in the case of artificial data, it is presented that the manual outlines dispersion is symmetric in relation to the truth border. The effectiveness of the selected segmentation algorithm was analysed in the context of kidney cancer using computed tomography data. For the calculated local inter-observer variability, 80.11% of the surface points generated by the kidney segmentation algorithm are within one expert outline standard deviation and 97.96% are within five. An error distribution shift in the direction of type I error equivalent was also observed. Finally, the significance of the local estimation of error type differences is presented. The article shows the greater usefulness and flexibility of the proposed framework in comparison to the state-of-the-art methods. The exemplary usage of the LEFMIS with or without inter-/intra-observer variability is also presented.

Automated Techniques for the Interpretation of Fetal Abnormalities: A Review

Ultrasound (US) image segmentation methods, focusing on techniques developed for fetal biometric parameters and nuchal translucency, are briefly reviewed. Ultrasound medical images can easily identify the fetus using segmentation techniques and calculate fetal parameters. It can timely find the fetal abnormality so that necessary action can be taken by the pregnant woman. Firstly, a detailed literature has been offered on fetal biometric parameters and nuchal translucency to highlight the investigation approaches with a degree of validation in diverse clinical domains. Then, a categorization of the bibliographic assessment of recent research effort in the segmentation field of ultrasound 2D fetal images has been presented. The fetal images of high-risk pregnant women have been taken into the routine and continuous monitoring of fetal parameters. These parameters are used for detection of fetal weight, fetal growth, gestational age, and any possible abnormality detection.

Evaluation of a novel software application for magnetic resonance distortion correction in cranial stereotactic radiosurgery

This study aimed to validate a novel commercially available software for correcting spatial distortion in cranial magnetic resonance (MR) images. This software has been used to assess the dosimetric impact of MR distortion in stereotactic radiosurgery (SRS) treatments of vestibular schwannomas (VSs). Five MR datasets were intentionally distorted. Each distorted MR dataset was corrected using the Cranial Distortion software, obtaining a new corrected MR dataset (MRcorr). The accuracy of the correction was quantified by calculating the target registration error (TRE) for 6 anatomical landmarks identified in the co-registered MRcorr and planning computed tomography (pCT) images. Nine VS cases were included to investigate the impact of the MR distortion in SRS plans. Each SRS plan was calculated on the pCT (1 × 1 × 1 mm³ voxel) with the target and organs at risk (OARs) delineated using the planning MR dataset. This MR dataset was then corrected (MRcorr) using the Cranial Distortion software. Geometrical agreement between the original target and the corresponding corrected target was assessed using several metrics: MacDonald criteria, mean distance to agreement (MDA), and Dice similarity coefficient (DSC). Target coverage (D99%) and maximum doses (D2%) to ipsilateral cochlea and brainstem resulting on the MRcorr dataset were compared with the original values. TRE values (0.6 mm ± 0.3 mm) and differences found in Macdonald criteria (0.3 mm ± 0.4 mm and 0.3 mm ± 0.3 mm) and MDA (0.8 mm ± 0.2 mm) were mostly within the voxel size dimension of the pCT scan (1 × 1 × 1 mm³). High similarity (DSC > 0.7) between the original and corrected targets was found. Small dose differences for the original and corrected structures were found: 0.1 Gy ± 0.1 Gy for target D99%, 0.2 Gy ± 0.3 Gy for cochlea D2%, and 0.1 Gy ± 0.1 Gy for brainstem D2%. Our study shows that Distortion Correction software can be a helpful tool to detect and adequately correct brain MR distortions. However, a negligible dosimetric impact of MR distortion has been detected in our clinical practice.

Automatic segmentation method of pelvic floor levator hiatus in ultrasound using a self-normalizing neural network

Segmentation of the levator hiatus in ultrasound allows the extraction of biometrics, which are of importance for pelvic floor disorder assessment. We present a fully automatic method using a convolutional neural network (CNN) to outline the levator hiatus in a two-dimensional image extracted from a three-dimensional ultrasound volume. In particular, our method uses a recently developed scaled exponential linear unit (SELU) as a nonlinear self-normalizing activation function, which for the first time has been applied in medical imaging with CNN. SELU has important advantages such as being parameter-free and mini-batch independent, which may help to overcome memory constraints during training. A dataset with 91 images from 35 patients during Valsalva, contraction, and rest, all labeled by three operators, is used for training and evaluation in a leave-one-patient-out cross validation. Results show a median Dice similarity coefficient of 0.90 with an interquartile range of 0.08, with equivalent performance to the three operators (with a Williams' index of 1.03), and outperforming a U-Net architecture without the need for batch normalization. We conclude that the proposed fully automatic method achieved equivalent accuracy in segmenting the pelvic floor levator hiatus compared to a previous semiautomatic approach.

Evaluation of localized region-based segmentation algorithms for CT-based delineation of organs at risk in radiotherapy

Background and purpose

In radiation therapy, defining the precise borders of cancerous tissues and adjacent normal organs has a significant effect on the therapy outcome. Deformable models offer a unique and robust approach to medical image segmentation. The objective of this study was to investigate the reliability of segmenting organs-at-risk (OARs) using three well-known local region-based level-set techniques.

Methods and materials

A total of 1340 non-enhanced and enhanced planning computed tomography (CT) slices of eight OARs (the bladder, rectum, kidney, clavicle, humeral head, femoral head, spinal cord, and lung) were segmented by using local region-based active contour, local Chan-Vese, and local Gaussian distribution models. Quantitative metrics, namely Hausdorff Distance (HD), Mean Absolute Distance (MAD), Dice coefficient (DC), Percentage Volume Difference (PVD) and Absolute Volumetric Difference (AVD), were adopted to measure the correspondence between detected contours and the manual references drawn by experts.

Results

The results showed the feasibility of using local region-based active contour methods for defining six of the OARs (the bladder, kidney, clavicle, humeral head, spinal cord, and lung) when adequate intensity information is available. While the most accurate results were achieved for lung (DC = 0.94) and humeral head (DC = 0.92), a poor level of agreement (DC < 0.7) was obtained for both rectum and femur.

Conclusion

Incorporating local statistical information in level set methods yields to satisfactory results of OARs delineation when adequate intensity information exists between the organs. However, the complexity of adjacent organs and the lack of distinct boundaries would result in a considerable segmentation error.

Dose accumulation of multiple high dose rate prostate brachytherapy treatments in two commercially available image registration systems

PURPOSE

The purpose of this study was to assess whether deformable image registration (DIR) is required for dose accumulation of multiple high dose rate prostate brachytherapy (HDRPBT) plans treated with the same catheter pattern on two different CT datasets.

METHOD

DIR was applied to 20 HDRPBT patients' planning CT images who received two treatment fractions on sequential days, on two different CT datasets, with the same implant. Quality of DIR in Velocity and MIM image registration systems was assessed by calculating the Dice Similarity Coefficient (DSC) and mean distance to agreement (MDA) for the prostate, urethra and rectum contours. Accumulated doses from each system were then calculated using the same DIR technique and dose volume histogram (DVH) parameters compared to manual addition with no DIR.

RESULTS

The average DSC was found to be 0.83 (Velocity) and 0.84 (MIM), 0.80 (Velocity) and 0.80 (MIM), 0.80 (Velocity) and 0.81 (MIM), for the prostate, rectum and urethra contours, respectively. The average difference in calculated DVH parameters between the two systems using dose accumulation was less than 1%, and there was no statistically significant difference found between deformably accumulated doses in the two systems versus manual DVH addition with no DIR.

CONCLUSION

Contour propagation using DIR in velocity and MIM was shown to be at least equivalent to inter-observer contouring variability on CT. The results also indicate that dose accumulation through manual addition of DVH parameters may be sufficient for HDRPBT treatments treated with the same catheter pattern on two different CT datasets.

ARCOCT: Automatic detection of lumen border in intravascular OCT images

BACKGROUND AND OBJECTIVE

Intravascular optical coherence tomography (OCT) is an invaluable tool for the detection of pathological features on the arterial wall and the investigation of post-stenting complications. Computational lumen border detection in OCT images is highly advantageous, since it may support rapid morphometric analysis. However, automatic detection is very challenging, since OCT images typically include various artifacts that impact image clarity, including features such as side branches and intraluminal blood presence. This paper presents ARCOCT, a segmentation method for fully-automatic detection of lumen border in OCT images.

METHODS

ARCOCT relies on multiple, consecutive processing steps, accounting for image preparation, contour extraction and refinement. In particular, for contour extraction ARCOCT employs the transformation of OCT images based on physical characteristics such as reflectivity and absorption of the tissue and, for contour refinement, local regression using weighted linear least squares and a 2nd degree polynomial model is employed to achieve artifact and small-branch correction as well as smoothness of the artery mesh. Our major focus was to achieve accurate contour delineation in the various types of OCT images, i.e., even in challenging cases with branches and artifacts.

RESULTS

ARCOCT has been assessed in a dataset of 1812 images (308 from stented and 1504 from native segments) obtained from 20 patients. ARCOCT was compared against ground-truth manual segmentation performed by experts on the basis of various geometric features (e.g. area, perimeter, radius, diameter, centroid, etc.) and closed contour matching indicators (the Dice index, the Hausdorff distance and the undirected average distance), using standard statistical analysis methods. The proposed method was proven very efficient and close to the ground-truth, exhibiting non statistically-significant differences for most of the examined metrics.

CONCLUSIONS

ARCOCT allows accurate and fully-automated lumen border detection in OCT images.