Statistical Region-Based Segmentation of Ultrasound Images

Segmentation of ultrasound images is a challenging problem due to speckle, which corrupts the image and can result in weak or missing image boundaries, poor signal to noise ratio and diminished contrast resolution. Speckle is a random interference pattern that is characterized by an asymmetric distribution as well as significant spatial correlation. These attributes of speckle are challenging to model in a segmentation approach, so many previous ultrasound segmentation methods simplify the problem by assuming that the speckle is white and/or Gaussian distributed. Unlike these methods, in this article we present an ultrasound-specific segmentation approach that addresses both the spatial correlation of the data as well as its intensity distribution. We first decorrelate the image and then apply a region-based active contour whose motion is derived from an appropriate parametric distribution for maximum likelihood image segmentation. We consider zero-mean complex Gaussian, Rayleigh, and Fisher-Tippett flows, which are designed to model fully formed speckle in the in-phase/quadrature (IQ), envelope detected, and display (log compressed) images, respectively. We present experimental results demonstrating the effectiveness of our method and compare the results with other parametric and nonparametric active contours. (E-mail:greg.slabaugh@gmail.com)     

Deep Learning-Based Carotid Plaque Segmentation from B-Mode Ultrasound Images

Carotid ultrasound measurement of total plaque area (TPA) provides a method for quantifying carotid plaque burden and monitoring changes in carotid atherosclerosis in response to medical treatment. Plaque boundary segmentation is required to generate the TPA measurement; however, training of observers and manual delineation are time consuming. Thus, our objective was to develop an automated plaque segmentation method to generate TPA from longitudinal carotid ultrasound images. In this study, a deep learning-based method, modified U-Net, was used to train the segmentation model and generate TPA measurement. A total of 510 plaques from 144 patients were used in our study, where the Monte Carlo cross-validation was used by randomly splitting the data set into 2/3 and 1/3 for training and testing. Two observers were trained to manually delineate the 510 plaques separately, which were used as the ground-truth references. Two U-Net models (M1 and M2) were trained using the two different ground-truth data sets from the two observers to evaluate the accuracy, variability and sensitivity on the ground-truth data sets used for training our method. The results of the algorithm segmentations of the two models yielded strong agreement with the two manual segmentations with the Pearson correlation coefficient r = 0.989 (p < 0.0001) and r = 0.987 (p < 0.0001). Comparison of the U-Net and manual segmentations resulted in mean TPA differences of 0.05 ± 7.13 mm² (95% confidence interval: 14.02–13.02 mm²) and 0.8 ± 8.7 mm² (17.85–16.25 mm²) for the two models, which are small compared with the TPA range in our data set from 4.7 to 312.8 mm². Furthermore, the mean time to segment a plaque was only 8.3 ± 3.1 ms. The presented deep learning-based method described has sufficient accuracy with a short computation time and exhibits high agreement between the algorithm and manual TPA measurements, suggesting that the method could be used to measure TPA and to monitor the progression and regression of carotid atherosclerosis.     

A Novel Model of Thyroid Nodule Segmentation for Ultrasound Images

Ultrasonography is regarded as an effective technique for the detection, diagnosis and monitoring of thyroid nodules. Segmentation of thyroid nodules on ultrasound images is important in clinical practice. However, because in ultrasound images there is an unclear boundary between thyroid nodules and surrounding tissues, the accuracy of segmentation remains a challenge. Although the deep learning model provides an accurate and convenient method for thyroid nodule segmentation, it is unsatisfactory of the existing model in segmenting the margin of thyroid nodules. In this study, we developed boundary attention transformer net (BTNet), a novel segmentation network with a boundary attention mechanism combining the advantages of a convolutional neural network and transformer, which could fuse the features of both long and short ranges. Boundary attention is improved to focus on learning the boundary information, and this module enhances the segmentation ability of the network boundary. For features of different scales, we also incorporate a deep supervision mechanism to blend the outputs of different levels to enhance the segmentation effect. As the BTNet model incorporates the long range–short range connectivity effect and the boundary–regional cooperation capability, our model has excellent segmentation performance in thyroid nodule segmentation. The development of BTNet was based on the data set from Shanghai Jiao Tong University School of Medicine Affiliated Sixth People's Hospital and the public data set. BTNet achieved good performance in the segmentation of thyroid nodules with an intersection-over-union of 0.810 and Dice coefficient of 0.892 Moreover, our work revealed great improvement in the boundary metrics; for example, the boundary distance was 7.308, the boundary overlap 0.201 and the boundary Dice 0.194, all with p values <0.05.     

Segmentation of Skin Tumors in High-Frequency 3-D Ultrasound Images

High-frequency 3-D ultrasound imaging is an informative tool for diagnosis, surgery planning and skin lesion examination. The purpose of this article was to describe a semi-automated segmentation tool providing easy access to the extent, shape and volume of a lesion. We propose an adaptive log-likelihood level-set segmentation procedure using non-parametric estimates of the intensity distribution. The algorithm has a single parameter to control the smoothness of the contour, and we describe how a fixed value yields satisfactory segmentation results with an average Dice coefficient of D = 0.76. The algorithm is implemented on a grid, which increases the speed by a factor of 100 compared with a standard pixelwise segmentation. We compare the method with parametric methods making the hypothesis of Rayleigh or Nakagami distributed signals, and illustrate that our method has greater robustness with similar computational speed. Benchmarks are made on realistic synthetic ultrasound images and a data set of nine clinical 3-D images acquired with a 50-MHz imaging system. The proposed algorithm is suitable for use in a clinical context as a post-processing tool.     

Fully Automatic Plaque Segmentation in 3-D Carotid Ultrasound Images

Automatic segmentation of the carotid plaques from ultrasound images has been shown to be an important task for monitoring progression and regression of carotid atherosclerosis. Considering the complex structure and heterogeneity of plaques, a fully automatic segmentation method based on media-adventitia and lumen-intima boundary priors is proposed. This method combines image intensity with structure information in both initialization and a level-set evolution process. Algorithm accuracy was examined on the common carotid artery part of 26 3-D carotid ultrasound images (34 plaques ranging in volume from 2.5 to 456 mm³) by comparing the results of our algorithm with manual segmentations of two experts. Evaluation results indicated that the algorithm yielded total plaque volume (TPV) differences of −5.3 ± 12.7 and −8.5 ± 13.8 mm³ and absolute TPV differences of 9.9 ± 9.5 and 11.8 ± 11.1 mm³. Moreover, high correlation coefficients in generating TPV (0.993 and 0.992) between algorithm results and both sets of manual results were obtained. The automatic method provides a reliable way to segment carotid plaque in 3-D ultrasound images and can be used in clinical practice to estimate plaque measurements for management of carotid atherosclerosis.     

Attention-Enriched Deep Learning Model for Breast Tumor Segmentation in Ultrasound Images

Incorporating human domain knowledge for breast tumor diagnosis is challenging because shape, boundary, curvature, intensity or other common medical priors vary significantly across patients and cannot be employed. This work proposes a new approach to integrating visual saliency into a deep learning model for breast tumor segmentation in ultrasound images. Visual saliency refers to image maps containing regions that are more likely to attract radiologists’ visual attention. The proposed approach introduces attention blocks into a U-Net architecture and learns feature representations that prioritize spatial regions with high saliency levels. The validation results indicate increased accuracy for tumor segmentation relative to models without salient attention layers. The approach achieved a Dice similarity coefficient (DSC) of 90.5% on a data set of 510 images. The salient attention model has the potential to enhance accuracy and robustness in processing medical images of other organs, by providing a means to incorporate task-specific knowledge into deep learning architectures.     

Segmentation of White Matter Flaring Areas in Ultrasound Images of Very-Low-Birth-Weight Preterm Infants

In this article, we present an interactive algorithm segmenting white brain matter, visible as hyperechoic flaring areas in ultrasound (US) images of preterm infants with periventricular leukomalacia (PVL). The algorithm combines both the textural properties of pathological brain tissue and mathematical morphology operations. An initial flaring area estimate is derived from a multifeature multiclassifier tissue texture classifier. This area is refined based on the structural properties of the choroid plexus, a brain feature known to have characteristics similar to flaring. Subsequently, a combination of a morphological closing, gradient and opening by reconstruction operation determines the final flaring area boundaries. Experimental results are compared with a gold standard constructed from manual flaring area delineations of 12 medical experts. In addition, we compared our algorithm to an existing active contour method. The results show our technique agrees to the gold standard with statistical significance and outperforms the existing method in accuracy. Finally, using the flaring area as a criterion we improve the sensitivity of PVL detection up to 98% as compared with the state of the art. (E-mail: ervsteen@telin.ugent.be)     

Automatic Segmentation of Calcifications in Intravascular Ultrasound Images Using Snakes and the Contourlet Transform

It is valuable to detect calcifications in intravascular ultrasound images for studies of coronary artery diseases. An image segmentation method based on snakes and the Contourlet transform is proposed to automatically and accurately detect calcifications. With the Contourlet transform, an original image is decomposed into low-pass bands and band-pass directional sub-bands. The 2-D Renyi's entropy is used to adaptively threshold the low-pass bands in a multiresolution hierarchy to determine regions-of-interest (ROIs). Then a mean intensity ratio, reflecting acoustic shadowing, is presented to classify calcifications from noncalcifications and obtain initial contours of calcifications. The anisotropic diffusion is used in bandpass directional sub-bands to suppress noise and preserve calcific edges. Finally, the contour deformation in the boundary vector field is used to obtain final contours of calcifications. The method was evaluated via 60 simulated images and 86 in vivo images. It outperformed a recently proposed method, the Santos Filho method, by 2.76% and 14.53%, in terms of the sensitivity and specificity of calcification detection, respectively. The area under the receiver operating characteristic curve increased by 0.041. The relative mean distance error, relative difference degree, relative arc difference, relative thickness difference and relative length difference were reduced by 5.73%, 19.79%, 11.62%, 12.06% and 20.51%, respectively. These results reveal that the proposed method can automatically and accurately detect calcifications and delineate their boundaries. (E-mail: yywang@fudan.edu.cn)     

User-Guided Segmentation of Preterm Neonate Ventricular System from 3-D Ultrasound Images Using Convex Optimization

A three-dimensional (3-D) ultrasound (US) system has been developed to monitor the intracranial ventricular system of preterm neonates with intraventricular hemorrhage (IVH) and the resultant dilation of the ventricles (ventriculomegaly). To measure ventricular volume from 3-D US images, a semi-automatic convex optimization-based approach is proposed for segmentation of the cerebral ventricular system in preterm neonates with IVH from 3-D US images. The proposed semi-automatic segmentation method makes use of the convex optimization technique supervised by user-initialized information. Experiments using 58 patient 3-D US images reveal that our proposed approach yielded a mean Dice similarity coefficient of 78.2% compared with the surfaces that were manually contoured, suggesting good agreement between these two segmentations. Additional metrics, the mean absolute distance of 0.65 mm and the maximum absolute distance of 3.2 mm, indicated small distance errors for a voxel spacing of 0.22 × 0.22 × 0.22 mm³. The Pearson correlation coefficient (r = 0.97, p < 0.001) indicated a significant correlation of algorithm-generated ventricular system volume (VSV) with the manually generated VSV. The calculated minimal detectable difference in ventricular volume change indicated that the proposed segmentation approach with 3-D US images is capable of detecting a VSV difference of 6.5 cm³ with 95% confidence, suggesting that this approach might be used for monitoring IVH patients' ventricular changes using 3-D US imaging. The mean segmentation times of the graphics processing unit (GPU)- and central processing unit-implemented algorithms were 50 ± 2 and 205 ± 5 s for one 3-D US image, respectively, in addition to 120 ± 10 s for initialization, less than the approximately 35 min required by manual segmentation. In addition, repeatability experiments indicated that the intra-observer variability ranges from 6.5% to 7.5%, and the inter-observer variability is 8.5% in terms of the coefficient of variation of the Dice similarity coefficient. The intra-class correlation coefficient for ventricular system volume measurements for each independent observer ranged from 0.988 to 0.996 and was 0.945 for three different observers. The coefficient of variation and intra-class correlation coefficient revealed that the intra- and inter-observer variability of the proposed approach introduced by the user initialization was small, indicating good reproducibility, independent of different users.     

医学超声图像的分辨力

B超仪的图像分辨力是仪器的主要性能参数之一,本文主要分析了影响超声图像的空间分辨力的各种因素。显示器一般都可显示500像素×500像素矩阵,其图像分辨力是足够好的。超声扫查线数量(通常≤256线)所决定的横向(侧向)空间采样分辨力在中远场区(对凸阵、相控阵)较差,电子聚焦系统的质量可能是决定B超仪横向分辨力的最主要的因素。成像系统的频带宽度是决定B超仪纵向(轴向)分辨力的关键因素。     

Deep Learning-Based Segmentation of Nodules in Thyroid Ultrasound: Improving Performance by Utilizing Markers Present in the Images

Computer-aided segmentation of thyroid nodules in ultrasound imaging could assist in their accurate characterization. In this study, using data for 1278 nodules, we proposed and evaluated two methods for deep learning-based segmentation of thyroid nodules that utilize calipers present in the images. The first method used approximate nodule masks generated based on the calipers. The second method combined manual annotations with automatic guidance by the calipers. When only approximate nodule masks were used for training, the achieved Dice similarity coefficient (DSC) was 85.1%. The performance of a network trained using manual annotations was DSC = 90.4%. When the guidance by the calipers was added, the performance increased to DSC = 93.1%. An increase in the number of cases used for training resulted in increased performance for all methods. The proposed method utilizing the guidance by calipers matched the performance of the network that did not use it with a reduced number of manually annotated training cases.     

Random Forest-Based Bone Segmentation in Ultrasound

Ultrasound (US) imaging is a safe alternative to radiography for guidance during minimally invasive orthopedic procedures. However, ultrasound is challenging to interpret because of the relatively low signal-to-noise ratio and its inherent speckle pattern that decreases image quality. Here we describe a method for automatic bone segmentation in 2-D ultrasound images using a patch-based random forest classifier and several ultrasound specific features, such as shadowing. We illustrate that existing shadow features are not robust to changes in US acquisition parameters, and propose a novel robust shadow feature. We evaluate the method on several US data sets and report that it favorably compares with existing techniques. We achieve a recall of 0.86 at a precision of 0.82 on a test set of 143 spinal US images.     

Segmentation of the Common Carotid Intima-Media Complex in Ultrasound Images Using 2-D Continuous Max-Flow and Stacked Sparse Auto-encoder

The intima-media thickness (IMT) of a common carotid artery in an ultrasound image is considered an important indicator of the onset of atherosclerosis. However, it is challenging to segment the intima-media complex (IMC) directly in ultrasound images. This study proposes a fully automatic method to segment the IMC on longitudinal B-mode ultrasound images. Our method consists of two stages: (i) extraction of the region of interest with a continuous max-flow algorithm and region-of-interest reconstruction using a stacked sparse auto-encoder model, and (ii) IMC segmentation using a trained random forest classifier. The proposed method has been tested on three databases from three different imaging centres, comprising a total of 228 ultrasound images of the common carotid artery. On the three databases, our method yields mean absolute errors of 0.028 ± 0.016 mm, 0.579 ± 0.288 pixel and 0.582 ± 0.341 pixel; polyline distance (PD) measures of 0.026 ± 0.017 mm, 0.657 ± 0.275 pixel and 0.731 ± 0:282 pixel; Hausdorff distance measures of 0.249 ± 0.101 mm, 4.760 ± 1.085 pixels and 5.825 ± 2.059 pixels; and correlation coefficients of 95.19%, 93.79%, and 98.96%, respectively. These results indicate that the proposed method performs well in segmentation of the IMC and measurement of the IMT.     

Speckles Suppression Techniques for Ultrasound Images

Most medical images have a poorer signal to noise ratio than scenes taken with a digital camera, which often leads to incorrect diagnosis. Speckles suppression from ultrasound images is one of the most important concerns in computer-aided diagnosis. This article proposes two novel, robust and efficient ultrasound images denoising techniques. The first technique is the enhanced ultrasound images denoising (EUID) technique, which estimates automatically the speckle noise amount in the ultrasound images by estimating important input parameters of the filter and then denoising the image using the sigma filter. The second technique is the ultrasound image denoising using neural network (UIDNN) that is based on the second-order difference of pixels with adaptive threshold value in order to identify random valued speckles from images to achieve high efficient image restoration. The performances of the proposed techniques are analyzed and compared with those of other image denoising techniques. The experimental results show that the proposed techniques are valuable tools for speckles suppression, being accurate, less tedious, and preventing typical human errors associated with manual tasks in addition to preserving the edges from the image. The EUID algorithm has nearly the same peak signal to noise ratio (PSNR) as Frost and speckle-reducing anisotropic diffusion 1, whereas it achieves higher gains, on average—0.4 dB higher PSNR—than the Lee, Kuan, and anisotropic diffusion filters. The UIDNN technique outperforms all the other techniques since it can determine the noisy pixels and perform filtering for these pixels only. Generally, when relatively high levels of noise are added, the proposed algorithms show better performances than the other conventional filters.     

Deconvolution of In Vivo Ultrasound B-Mode Images

An algorithm for deconvolution of medical ultrasound images is presented. The procedure involves estimation of the basic one-dimensional ultrasound pulse, determining the ratio of the covariance of the noise to the covariance of the reflection signal, and finally deconvolution of the rf signal from the transducer. Using pulse and covariance estimators makes the approach self-calibrating, as all parameters for the procedure are estimated from the patient under investigation. An example of use on a clinical, in-vivo image is given. A 2 × 2 cm region of the portal vein in a liver is deconvolved. An increase in axial resolution by a factor of 2.4 is obtained. The procedure can also be applied to whole images, when it is ensured that the rf signal is properly measured. A method for doing that is outlined.     

Visual Screening of Muscle Ultrasound Images in Children

In children, non-invasive muscle ultrasound (MU) imaging has become increasingly important for the detection of neuromuscular pathology, by either quantitative or visual assessment. MU quantification requires time, expertise and equipment. If application of visual MU screening provides reliable results, ubiquitous application could be advocated. Previously, we found that visual MU screening can reliably detect segmental neuromuscular alterations within a patient. Analogously, we reasoned that visual MU screening could discern pathologic MU images from healthy controls. We therefore investigated visual screening results by 20 clinical observers (involving 100 MU images, with [n = 53] and without [n = 47] neuromuscular pathology). MU screening revealed adequate sensitivity, specificity and negative predictive value (85%, 75% and 82%, respectively). MU-experienced observers revealed higher specificity than MU-inexperienced observers (86% vs. 69%, p = 0.005). We conclude that clinical observers can identify neuromuscular pathology by visual screening. To enhance specificity, a secondary view by an expert appears advisory.     

Diagnosis of Metacarpophalangeal Synovitis with Musculoskeletal Ultrasound Images

Rheumatoid arthritis (RA) is a chronic autoimmune disease that can result in considerable disability and pain. The metacarpophalangeal (MCP) joint is the most common diseased joint in RA. In clinical practice, MCP synovitis is commonly diagnosed on the basis of musculoskeletal ultrasound (MSUS) images. However, because of the vague criteria, the consistency in grading MCP synovitis based on MSUS images fluctuates between ultrasound imaging practitioners. Therefore, a new method for diagnosis of MCP synovitis is needed. Deep learning has developed rapidly in the medical area, which often requires a large-scale data set. However, the total number of MCP-MSUS images fell far short of the demand, and the distribution of different medical grades of images was unbalanced. With use of the traditional image augmentation methods, the diversity of the data remains insufficient. In this study, a high-resolution generative adversarial network (HRGAN) method that generates enough images for network training and enriches the diversity of the training data set is described. In comparison experiments, our proposed diagnostic system based on MSUS images provided more consistent results than those provided by clinical physicians. As the proposed method is image relevant, this study might provide a reference for other medical image classification research with insufficient data sets.     

Segmentation of B-mode cardiac ultrasound data by Bayesian Probability Maps

In this paper we present a model for describing the position distribution of the endocardium in the two-chamber apical long-axis view of the heart in clinical B-mode ultrasound cycles. We propose a novel Bayesian formulation, including priors for spatial and temporal smoothness, and preferred shapes and position. The shape model takes into account both endocardium, atrial region and apex. The likelihood is built using a statistical signal model, which attempts to closely model a censored signal. In addition, the use of a censored Gamma mixture model with unknown censoring point, to handle artefacts resulting from left-censoring of the in US clinical B-mode, is to our knowledge novel. The posterior density is sampled by the Gibbs method to estimate the expected latent variable representation of the endocardium, which we call the Bayesian Probability Map; the map describes the probability of pixels being classified as being within the endocardium. The regularization parameters of the model are estimated by cross-validation, and the results are compared against the two-chamber apical model of Chen et al.