Average DSC Research Articles

Randomized multi-reader evaluation of automated detection and segmentation of brain tumors in stereotactic radiosurgery with deep neural networks.

BackgroundStereotactic radiosurgery (SRS), a validated treatment for brain tumors, requires accurate tumor contouring. This manual segmentation process is time-consuming and prone to substantial inter-practitioner variability. Artificial intelligence (AI) with deep neural networks have increasingly been proposed for use in lesion detection and segmentation but have seldom been validated in a clinical setting.MethodsWe conducted a randomized, cross-modal, multi-reader, multispecialty, multi-case study to evaluate the impact of AI assistance on brain tumor SRS. A state-of-the-art auto-contouring algorithm built on multi-modality imaging and ensemble neural networks was integrated into the clinical workflow. Nine medical professionals contoured the same case series in two reader modes (assisted or unassisted) with a memory washout period of 6 weeks between each section. The case series consisted of 10 algorithm-unseen cases, including five cases of brain metastases, three of meningiomas, and two of acoustic neuromas. Among the nine readers, three experienced experts determined the ground truths of tumor contours.ResultsWith the AI assistance, the inter-reader agreement significantly increased (Dice similarity coefficient [DSC] from 0.86 to 0.90, P < 0.001). Algorithm-assisted physicians demonstrated a higher sensitivity for lesion detection than unassisted physicians (91.3% vs 82.6%, P = .030). AI assistance improved contouring accuracy, with an average increase in DSC of 0.028, especially for physicians with less SRS experience (average DSC from 0.847 to 0.865, P = .002). In addition, AI assistance improved efficiency with a median of 30.8% time-saving. Less-experienced clinicians gained prominent improvement on contouring accuracy but less benefit in reduction of working hours. By contrast, SRS specialists had a relatively minor advantage in DSC, but greater time-saving with the aid of AI.ConclusionsDeep learning neural networks can be optimally utilized to improve accuracy and efficiency for the clinical workflow in brain tumor SRS.

Automatic segmentation of white matter hyperintensities from brain magnetic resonance images in the era of deep learning and big data - A systematic review.

White matter hyperintensities (WMH), of presumed vascular origin, are visible and quantifiable neuroradiological markers of brain parenchymal change. These changes may range from damage secondary to inflammation and other neurological conditions, through to healthy ageing. Fully automatic WMH quantification methods are promising, but still, traditional semi-automatic methods seem to be preferred in clinical research. We systematically reviewed the literature for fully automatic methods developed in the last five years, to assess what are considered state-of-the-art techniques, as well as trends in the analysis of WMH of presumed vascular origin. We registered the systematic review protocol with the International Prospective Register of Systematic Reviews (PROSPERO), registration number - CRD42019132200. We conducted the search for fully automatic methods developed from 2015 to July 2020 on Medline, Science direct, IEE Explore, and Web of Science. We assessed risk of bias and applicability of the studies using QUADAS 2. The search yielded 2327 papers after removing 104 duplicates. After screening titles, abstracts and full text, 37 were selected for detailed analysis. Of these, 16 proposed a supervised segmentation method, 10 proposed an unsupervised segmentation method, and 11 proposed a deep learning segmentation method. Average DSC values ranged from 0.538 to 0.91, being the highest value obtained from an unsupervised segmentation method. Only four studies validated their method in longitudinal samples, and eight performed an additional validation using clinical parameters. Only 8/37 studies made available their methods in public repositories. We found no evidence that favours deep learning methods over the more established k-NN, linear regression and unsupervised methods in this task. Data and code availability, bias in study design and ground truth generation influence the wider validation and applicability of these methods in clinical research.

Automatic Microplot Localization Using UAV Images and a Hierarchical Image-Based Optimization Method.

To develop new crop varieties and monitor plant growth, health, and traits, automated analysis of aerial crop images is an attractive alternative to time-consuming manual inspection. To perform per-microplot phenotypic analysis, localizing and detecting individual microplots in an orthomosaic image of a field are major steps. Our algorithm uses an automatic initialization of the known field layout over the orthomosaic images in roughly the right position. Since the orthomosaic images are stitched from a large number of smaller images, there can be distortion causing microplot rows not to be entirely straight and the automatic initialization to not correctly position every microplot. To overcome this, we have developed a three-level hierarchical optimization method. First, the initial bounding box position is optimized using an objective function that maximizes the level of vegetation inside the area. Then, columns of microplots are repositioned, constrained by their expected spacing. Finally, the position of microplots is adjusted individually using an objective function that simultaneously maximizes the area of the microplot overlapping vegetation, minimizes spacing variance between microplots, and maximizes each microplot's alignment relative to other microplots in the same row and column. The orthomosaics used in this study were obtained from multiple dates of canola and wheat breeding trials. The algorithm was able to detect 99.7% of microplots for canola and 99% for wheat. The automatically segmented microplots were compared to ground truth segmentations, resulting in an average DSC of 91.2% and 89.6% across all microplots and orthomosaics in the canola and wheat datasets.

Brain tumor segmentation using 3D Mask R-CNN for dynamic susceptibility contrast enhanced perfusion imaging

The segmentation of neoplasms is an important part of radiotherapy treatment planning, monitoring disease progression, and predicting patient outcome. In the brain, functional magnetic resonance imaging (MRI) like dynamic susceptibility contrast enhanced (DSCE) or T1-weighted dynamic contrast enhanced (DCE) perfusion MRI are important tools for diagnosis. They play a crucial role in providing pre-operative assessment of tumor histology, grading, and biopsy guidance. However, the manual contouring of these neoplasms is tedious, expensive, time-consuming, and vulnerable to inter-observer variability. In this work, we propose a 3D mask region-based convolutional neural network (R-CNN) method to automatically segment brain tumors in DSCE MRI perfusion images. As our goal is to simultaneously localize and segment the tumor, our training process contained both a region-of-interest (ROI) localization and regression with voxel-wise segmentation. The combination of classification loss, ROI location and size regression loss, and segmentation loss were used to supervise the proposed network. We retrospectively investigated 21 patients’ perfusion images, with between 50 and 70 perfusion time point volumes, a total of 1260 3D volumes. Tumor contours were automatically segmented by our proposed method and compared against other state-of-the-art methods and those delineated by physicians as the ground truth. The results of our method demonstrated good agreement with the ground truth contours. The average DSC, precision, recall, Hausdorff distance, mean surface distance (MSD), root MSD, and center of mass distance were 0.90 ± 0.04, 0.91 ± 0.04, 0.90 ± 0.06, 7.16 ± 5.78 mm, 0.45 ± 0.34 mm, 1.03 ± 0.72 mm, and 0.86 ± 0.91 mm, respectively. These results support the feasibility of our method in accurately localizing and segmenting brain tumors in DSCE perfusion MRI. Our 3D Mask R-CNN segmentation method in DSCE perfusion imaging has great promise for future clinical use.

Pathological lung segmentation based on random forest combined with deep model and multi-scale superpixels.

Accurate segmentation of lungs in pathological thoracic computed tomography (CT) scans plays an important role in pulmonary disease diagnosis. However, it is still a challenging task due to the variability of pathological lung appearances and shapes. In this paper, we proposed a novel segmentation algorithm based on random forest (RF), deep convolutional network, and multi-scale superpixels for segmenting pathological lungs from thoracic CT images accurately. A pathological thoracic CT image is first segmented based on multi-scale superpixels, and deep features, texture, and intensity features extracted from superpixels are taken as inputs of a group of RF classifiers. With the fusion of classification results of RFs by a fractional-order gray correlation approach, we capture an initial segmentation of pathological lungs. We finally utilize a divide-and-conquer strategy to deal with segmentation refinement combining contour correction of left lungs and region repairing of right lungs. Our algorithm is tested on a group of thoracic CT images affected with interstitial lung diseases. Experiments show that our algorithm can achieve a high segmentation accuracy with an average DSC of 96.45% and PPV of 95.07%. Compared with several existing lung segmentation methods, our algorithm exhibits a robust performance on pathological lung segmentation. Our algorithm can be employed reliably for lung field segmentation of pathologic thoracic CT images with a high accuracy, which is helpful to assist radiologists to detect the presence of pulmonary diseases and quantify its shape and size in regular clinical practices.

PSIGAN: Joint Probabilistic Segmentation and Image Distribution Matching for Unpaired Cross-Modality Adaptation-Based MRI Segmentation.

We developed a new joint probabilistic segmentation and image distribution matching generative adversarial network (PSIGAN) for unsupervised domain adaptation (UDA) and multi-organ segmentation from magnetic resonance (MRI) images. Our UDA approach models the co-dependency between images and their segmentation as a joint probability distribution using a new structure discriminator. The structure discriminator computes structure of interest focused adversarial loss by combining the generated pseudo MRI with probabilistic segmentations produced by a simultaneously trained segmentation sub-network. The segmentation sub-network is trained using the pseudo MRI produced by the generator sub-network. This leads to a cyclical optimization of both the generator and segmentation sub-networks that are jointly trained as part of an end-to-end network. Extensive experiments and comparisons against multiple state-of-the-art methods were done on four different MRI sequences totalling 257 scans for generating multi-organ and tumor segmentation. The experiments included, (a) 20 T1-weighted (T1w) in-phase mdixon and (b) 20 T2-weighted (T2w) abdominal MRI for segmenting liver, spleen, left and right kidneys, (c) 162 T2-weighted fat suppressed head and neck MRI (T2wFS) for parotid gland segmentation, and (d) 75 T2w MRI for lung tumor segmentation. Our method achieved an overall average DSC of 0.87 on T1w and 0.90 on T2w for the abdominal organs, 0.82 on T2wFS for the parotid glands, and 0.77 on T2w MRI for lung tumors.

A learning-based automatic segmentation and quantification method on left ventricle in gated myocardial perfusion SPECT imaging: A feasibility study

BackgroundThe performance of left ventricular (LV) functional assessment using gated myocardial perfusion SPECT (MPS) relies on the accuracy of segmentation. Current methods require manual adjustments that are tedious and subjective. We propose a novel machine-learning-based method to automatically segment LV myocardium and measure its volume in gated MPS imaging without human intervention. MethodsWe used an end-to-end fully convolutional neural network to segment LV myocardium by delineating its endocardial and epicardial surface. A novel compound loss function, which encourages similarity and penalizes discrepancy between prediction and training dataset, is utilized in training stage to achieve excellent performance. We retrospectively investigated 32 normal patients and 24 abnormal patients, whose LV myocardial contours automatically segmented by our method were compared with those delineated by physicians as the ground truth. ResultsThe results of our method demonstrated very good agreement with the ground truth. The average DSC metrics and Hausdorff distance of the contours delineated by our method are larger than 0.900 and less than 1 cm, respectively, among all 32 + 24 patients of all phases. The correlation coefficient of the LV myocardium volume between ground truth and our results is 0.910 ± 0.061 (P < 0.001), and the mean relative error of LV myocardium volume is − 1.09 ± 3.66%. ConclusionThese results strongly indicate the feasibility of our method in accurately quantifying LV myocardium volume change over the cardiac cycle. The learning-based segmentation method in gated MPS imaging has great promise for clinical use.

Improving brain tumor segmentation on MRI based on the deep U-net and residual units.

BACKGROUNDAccurate segmentation of brain tumor depicting on magnetic resonance imaging (MRI) is an important step for doctors to determine optimal treatment plan of Gliomas, which are the common malignant brain tumors that seriously damage patients' health and life.OBJECTThis study aims to improve accuracy and efficiency of brain tumor segmentation on MRI using the advanced deep learning model.METHODIn this study, an improved model based on the U-net for accurate segmentation of brain tumor MRI images, called Deeper ResU-net, is proposed. First, a deep Deeper U-net is built, which has deeper network depth compared with U-net, uses Squeeze Operator to control network parameters and attempts to enhance the feature extraction ability. Then, Deeper ResU-net is formed to eliminate degradation phenomenon of the deep network, in which residual unit is designed and integrated into the Deeper U-net to keep the number of parameters unchanged.RESULTDeeper ResU-net makes the deep network conduct stable training without degrading. Evaluation result shows that the Deeper ResU-net has achieved competitive result with average DSC metrics of 0.9, 0.82, 0.88 for Complete tumor region, Core tumor region and Enhanced tumor region, respectively.CONCLUSIONBy extending the U-net model to a deeper layer and adding the residual structure to ensure effective and stable training of the model, the experiment results demonstrate that applying the improved Deeper ResU-net can effectively eliminate the degradation phenomenon of deep network and improve segmentation performance.

Automatic segmentation of organs at risk for nasopharyngeal carcinoma with Smart Segmentation and MIM Atlas

Objective To compare the accuracy of two automatic segmentation softwares (Smart Segmentation and MIM Atlas) in organs at risk (OARs) contouring for nasopharyngeal carcinoma (NPC). Methods Totally 55 NPC patients were retrospectively reviewed with manually contoured OARs on CT images, in which 30 cases were randomly selected to create a data base in the Smart Segmentation and MIM Atlas. The remaining 25 cases were automatically contoured with Smart Segmentation and MIM as test cases. The automatic contouring accuracies of two softwares were evaluated with Dice coefficient(DSC), Hausdorff distance(HD), and absolute volume difference(△V) using manual contours as a golden standard. Results The overall DSC, HD and △V of all organs contoured by MIM Atlas and Smart Segmentation were (0.79±0.13) vs. (0.62±0.24) (t=14.06, P<0.05), (5.50±3.84)mm vs.(8.38±4.88)mm (t=-11.40, P<0.05), and (1.52±2.46) cm3vs. (2.38±3.57) cm3 (t=-4.70, P<0.05), respectively. The average DSC of 11 organs (brain stem, optic chiasm, bilateral lens, bilateral optic nerve, bilateral eyeballs, bilateral parotid gland, spinal cord) delineated by MIM Atlas was statistically greater than that of Smart Segmentation (t=5.27, 4.41, 6.34, 5.70, 10.62, 7.45, 3.96, 4.26, 6.25, 5.42, 7.23, P<0.05). The average HD of 10 organs (brain stem, optic chiasm, bilateral lens, bilateral optic nerve, bilateral eyeballs, left parotid gland, spinal cord) delineated by MIM Atlas was statistically less than that of Smart Segmentation (t=-4.51, -4.49, -3.92, -3.45, -5.36, -5.56, -3.89, -3.90, -3.60, -3.68, P<0.05). The average △V of 6 organs (brain stem, optic chiasm, left len, bilateral optic nerve, right eyeball) delineated by MIM Atlas was statistically less than that of Smart Segmentation (t=-2.83, -3.39, -2.56, -2.27, -2.43, -2.51, P<0.05). Conclusions Both softwares have reasonable contouring accuracy for larger organs. The accuracy decreased with the decrease of organ volumes and blurred boundary. Generally, MIM Atlas′s performs better than Smart Segmentation does. Key words: Automatic contouring; Organs-at-risk segmentation; Atlas library; Nasopharyngeal carcinoma

Carotid artery segmentation using level set method with double adaptive threshold (DATLS) on TOF-MRA images

PurposeThis study aimed to propose a method to semi-automatically segment lumen of carotid artery on TOF-MRA images with well performance on images with weak boundary. MethodsThe proposed method modified traditional level set method with double adaptive threshold (DATLS). Two thresholds were calculated from probability density function of initial region of interest. Threshold Th1 was used to generate initial contour and threshold Th2 was used to control evolution by modifying energy function of the level set. The performance of the proposed method was tested on TOF-MRA images from CARE II study via using manual delineation as reference. Sixty cases were randomly selected to compare the performance of traditional level set with DATLS, all 283 cases were used to test the robustness of DATLS and 20 cases were used to compare the intra- and inter-operator reproducibility of manual delineation with DATLS. The Dice Similarity Coefficient, Mean Contour Distance and Hausdorff Distance between the proposed method and the manual segmentation were reported. ResultsTraditional level set failed to segment carotid artery on 13 cases because of over convergence. Compared to traditional level set, DATLS showed a higher DSC (0.88 ± 0.07 vs. 0.75 ± 0.15), lower MCD (0.48 ± 0.37 mm vs. 1.89 ± 2.24 mm) and HD (1.41 ± 1.11 mm vs. 4.59 ± 4.59 mm) in the remaining 47 cases. The average DSC, MCD and HD of DATLS on 283 cases were 0.87 ± 0.09, 0.64 ± 0.87 mm and 1.76 ± 2.23 mm, respectively. Compared to manual delineation, DATLS had better reproducibility on DSC (intra-operator: 0.97 ± 0.09 and inter-operator: 0.97 ± 0.09 vs. intra-operator: 0.91 ± 0.04 and inter-operator: 0.91 ± 0.04), as well as MCD (intra-operator: 0.08 ± 0.18 mm and inter-operator: 0.13 ± 0.52 mm vs. intra-operator: 0.35 ± 0.20 mm and inter-operator: 0.37 ± 0.19 mm) and HD (intra-operator: 0.30 ± 0.66 mm and inter-operator: 0.48 ± 1.38 mm vs. intra-operator: 0.99 ± 0.79 mm and inter-operator: 0.90 ± 0.85 mm). ConclusionCompared to traditional level set, DATLS out-performed on computing time, robustness and accuracy of segmentation on TOF-MRA. The proposed method using modified level set with double adaptive threshold might be a promising tool for lumen segmentation of carotid artery on TOF-MRA.

A Deep Learning Approach to Vascular Structure Segmentation in Dermoscopy Colour Images.

Background Atypical vascular pattern is one of the most important features by differentiating between benign and malignant pigmented skin lesions. Detection and analysis of vascular structures is a necessary initial step for skin mole assessment; it is a prerequisite step to provide an accurate outcome for the widely used 7-point checklist diagnostic algorithm. Methods In this research we present a fully automated machine learning approach for segmenting vascular structures in dermoscopy colour images. The U-Net architecture is based on convolutional networks and designed for fast and precise segmentation of images. After preprocessing the images are randomly divided into 146516 patches of 64 × 64 pixels each. Results On the independent validation dataset including 74 images our implemented method showed high segmentation accuracy. For the U-Net convolutional neural network, an average DSC of 0.84, sensitivity 0.85, and specificity 0.81 has been achieved. Conclusion Vascular structures due to small size and similarity to other local structures create enormous difficulties during the segmentation and assessment process. The use of advanced segmentation methods like deep learning, especially convolutional neural networks, has the potential to improve the accuracy of advanced local structure detection.

Practical quantification of image registration accuracy following the AAPM TG-132 report framework.

The AAPM TG 132 Report enumerates important steps for validation of the medical image registration process. While the Report outlines the general goals and criteria for the tests, specific implementation may be obscure to the wider clinical audience. We endeavored to provide a detailed step‐by‐step description of the quantitative tests’ execution, applied as an example to a commercial software package (Mirada Medical, Oxford, UK), while striving for simplicity and utilization of readily available software. We demonstrated how the rigid registration data could be easily extracted from the DICOM registration object and used, following some simple matrix math, to quantify accuracy of rigid translations and rotations. The options for validating deformable image registration (DIR) were enumerated, and it was shown that the most practically viable ones are comparison of propagated internal landmark points on the published datasets, or of segmented contours that can be generated locally. The multimodal rigid registration in our example did not always result in the desired registration error below ½ voxel size, but was considered acceptable with the maximum errors under 1.3 mm and 1°. The DIR target registration errors in the thorax based on internal landmarks were far in excess of the Report recommendations of 2 mm average and 5 mm maximum. On the other hand, evaluation of the DIR major organs’ contours propagation demonstrated good agreement for lung and abdomen (Dice Similarity Coefficients, DSC, averaged over all cases and structures of 0.92 ± 0.05 and 0.91 ± 0.06, respectively), and fair agreement for Head and Neck (average DSC = 0.73 ± 0.14). The average for head and neck is reduced by small volume structures such as pharyngeal constrictor muscles. Even these relatively simple tests show that commercial registration algorithms cannot be automatically assumed sufficiently accurate for all applications. Formalized task‐specific accuracy quantification should be expected from the vendors.

Automated Segmentation Methods of Drusen to Diagnose Age-Related Macular Degeneration Screening in Retinal Images.

Existing drusen measurement is difficult to use in clinic because it requires a lot of time and effort for visual inspection. In order to resolve this problem, we propose an automatic drusen detection method to help clinical diagnosis of age-related macular degeneration. First, we changed the fundus image to a green channel and extracted the ROI of the macular area based on the optic disk. Next, we detected the candidate group using the difference image of the median filter within the ROI. We also segmented vessels and removed them from the image. Finally, we detected the drusen through Renyi's entropy threshold algorithm. We performed comparisons and statistical analysis between the manual detection results and automatic detection results for 30 cases in order to verify validity. As a result, the average sensitivity was 93.37% (80.95%~100%) and the average DSC was 0.73 (0.3~0.98). In addition, the value of the ICC was 0.984 (CI: 0.967~0.993, p < 0.01), showing the high reliability of the proposed automatic method. We expect that the automatic drusen detection helps clinicians to improve the diagnostic performance in the detection of drusen on fundus image.

Ground-glass nodule segmentation in chest CT images using asymmetric multi-phase deformable model and pulmonary vessel removal

We propose a ground-glass nodule (GGN) segmentation method that can separate solid component and ground-glass opacity (GGO) using an asymmetric multi-phase deformable model in chest CT images. First, initial solid component and GGO were extracted using intensity-based segmentation with histogram modeling. Second, the initial extracted regions were refined using an asymmetric multi-phase deformable model with modified energy functional and intensity-constrained averaging function. Finally, vessel-like structures are removed based on multi-scale shape analysis. In experiments, the segmentation accuracy of the entire GGN was evaluated using datasets from SNUH and LIDC/IDRI. The average DSC values of Seoul National University Hospital (SNUH) and Lung Image Database Consortium and Image Database Resource Initiative (LIDC/IDRI) were 0.85 ± 0.05 and 0.78 ± 0.07, respectively. The Pearson's correlation coefficient (r) between segmented volumes by the proposed method and manual segmentation was evaluated using SNUH dataset. The r of solid component, GGO, and entire GGN were 0.931, 0.875 and 0.907. Our experimental results show that the proposed method improves segmentation accuracy by applying the proposed asymmetric multiphase deformable model and pulmonary vessel removal.

Automatic Lumen Segmentation in Intravascular Optical Coherence Tomography Images Using Level Set.

Automatic lumen segmentation from intravascular optical coherence tomography (IVOCT) images is an important and fundamental work for diagnosis and treatment of coronary artery disease. However, it is a very challenging task due to irregular lumen caused by unstable plaque and bifurcation vessel, guide wire shadow, and blood artifacts. To address these problems, this paper presents a novel automatic level set based segmentation algorithm which is very competent for irregular lumen challenge. Before applying the level set model, a narrow image smooth filter is proposed to reduce the effect of artifacts and prevent the leakage of level set meanwhile. Moreover, a divide-and-conquer strategy is proposed to deal with the guide wire shadow. With our proposed method, the influence of irregular lumen, guide wire shadow, and blood artifacts can be appreciably reduced. Finally, the experimental results showed that the proposed method is robust and accurate by evaluating 880 images from 5 different patients and the average DSC value was 98.1% ± 1.1%.

복부 컴퓨터단층촬영 영상에서 다중 아틀라스 기반 위치적 정보를 사용한 계층적 장기 분할

In this paper, we propose an automatic hierarchical organ segmentation method on abdominal CT images. First, similar atlases are selected using bone-based similarity registration and similarity of liver, kidney, and pancreas area. Second, each abdominal organ is roughly segmented using image-based similarity registration and intensity-based locally weighted voting. Finally, the segmented abdominal organ is refined using mask-based affine registration and intensity-based locally weighted voting. Especially, gallbladder and pancreas are hierarchically refined using location information of neighbor organs such as liver, left kidney and spleen. Our method was tested on a dataset of 12 portal-venous phase CT data. The average DSC of total organs was <TEX>$90.47{\pm}1.70%$</TEX>. Our method can be used for patient-specific abdominal organ segmentation for rehearsal of laparoscopic surgery.

Temporal Feature Extraction from DCE-MRI to Identify Poorly Perfused Subvolumes of Tumors Related to Outcomes of Radiation Therapy in Head and Neck Cancer.

This study aimed to develop an automated model to extract temporal features from DCE-MRI in head-and-neck (HN) cancers to localize significant tumor subvolumes having low blood volume (LBV) for predicting local and regional failure after chemoradiation therapy. Temporal features were extracted from time-intensity curves to build classification model for differentiating voxels with LBV from those with high BV. Support vector machine (SVM) classification was trained on the extracted features for voxel classification. Subvolumes with LBV were then assembled from the classified voxels with LBV. The model was trained and validated on independent datasets created from 456 873 DCE curves. The resultant subvolumes were compared to ones derived by a 2-step method via pharmacokinetic modeling of blood volume, and evaluated for classification accuracy and volumetric similarity by DSC. The proposed model achieved an average voxel-level classification accuracy and DSC of 82% and 0.72, respectively. Also, the model showed tolerance on different acquisition parameters of DCE-MRI. The model could be directly used for outcome prediction and therapy assessment in radiation therapy of HN cancers, or even supporting boost target definition in adaptive clinical trials with further validation. The model is fully automatable, extendable, and scalable to extract temporal features of DCE-MRI in other tumors.

Combining Population and Patient-Specific Characteristics for Prostate Segmentation on 3D CT Images.

Prostate segmentation on CT images is a challenging task. In this paper, we explore the population and patient-specific characteristics for the segmentation of the prostate on CT images. Because population learning does not consider the inter-patient variations and because patient-specific learning may not perform well for different patients, we are combining the population and patient-specific information to improve segmentation performance. Specifically, we train a population model based on the population data and train a patient-specific model based on the manual segmentation on three slice of the new patient. We compute the similarity between the two models to explore the influence of applicable population knowledge on the specific patient. By combining the patient-specific knowledge with the influence, we can capture the population and patient-specific characteristics to calculate the probability of a pixel belonging to the prostate. Finally, we smooth the prostate surface according to the prostate-density value of the pixels in the distance transform image. We conducted the leave-one-out validation experiments on a set of CT volumes from 15 patients. Manual segmentation results from a radiologist serve as the gold standard for the evaluation. Experimental results show that our method achieved an average DSC of 85.1% as compared to the manual segmentation gold standard. This method outperformed the population learning method and the patient-specific learning approach alone. The CT segmentation method can have various applications in prostate cancer diagnosis and therapy.

Optimal number of atlases and label fusion for automatic multi-atlas-based brachial plexus contouring in radiotherapy treatment planning.

BackgroundThe present study aimed to define the optimal number of atlases for automatic multi-atlas-based brachial plexus (BP) segmentation and to compare Simultaneous Truth and Performance Level Estimation (STAPLE) label fusion with Patch label fusion using the ADMIRE® software. The accuracy of the autosegmentations was measured by comparing all of the generated autosegmentations with the anatomically validated gold standard segmentations that were developed using cadavers.Materials and methodsTwelve cadaver computed tomography (CT) atlases were used for automatic multi-atlas-based segmentation. To determine the optimal number of atlases, one atlas was selected as a patient and the 11 remaining atlases were registered onto this patient using a deformable image registration algorithm. Next, label fusion was performed by using every possible combination of 2 to 11 atlases, once using STAPLE and once using Patch. This procedure was repeated for every atlas as a patient.The similarity of the generated automatic BP segmentations and the gold standard segmentation was measured by calculating the average Dice similarity (DSC), Jaccard (JI) and True positive rate (TPR) for each number of atlases. These similarity indices were compared for the different number of atlases using an equivalence trial and for the two label fusion groups using an independent sample-t test.ResultsDSC’s and JI’s were highest when using nine atlases with both STAPLE (average DSC = 0,532; JI = 0,369) and Patch (average DSC = 0,530; JI = 0,370). When comparing both label fusion algorithms using 9 atlases for both, DSC and JI values were not significantly different. However, significantly higher TPR values were achieved in favour of STAPLE (p < 0,001). When fewer than four atlases were used, STAPLE produced significantly lower DSC, JI and TPR values than did Patch (p = 0,0048).ConclusionsUsing 9 atlases with STAPLE label fusion resulted in the most accurate BP autosegmentations (average DSC = 0,532; JI = 0,369 and TPR = 0,760). Only when using fewer than four atlases did the Patch label fusion results in a significantly more accurate autosegmentation than STAPLE.

Primary lung tumor segmentation from PET-CT volumes with spatial-topological constraint.

Accurate lung tumor segmentation is a prerequisite for effective radiation therapy and surgical planning. However, tumor delineation is challenging when the tumor boundaries are indistinct on PET or CT. To address this problem, we developed a segmentation method to improve the delineation of primary lung tumors from PET-CT images. We formulated the segmentation problem as a label information propagation process in an iterative manner. Our model incorporates spatial-topological information from PET and local intensity changes from CT. The topological information of the regions was extracted based on the metabolic activity of different tissues. The spatial-topological information moderates the amount of label information that a pixel receives: The label information attenuates as the spatial distance increases and when crossing different topological regions. Thus, the spatial-topological constraint assists accurate tumor delineation and separation. The label information propagation and transition model are solved under a random walk framework. Our method achieved an average DSC of 0.848 ± 0.036 and HD (mm) of 8.652 ± 4.532 on 40 patients with lung cancer. The t test showed a significant improvement (p value < 0.05) in segmentation accuracy when compared to eight other methods. Our method was better able to delineate tumors that had heterogeneous FDG uptake and which abutted adjacent structures that had similar densities. Our method, using a spatial-topological constraint, provided better lung tumor delineation, in particular, when the tumor involved or abutted the chest wall and the mediastinum.