Comparing deep learning-based automatic segmentation of breast masses to expert interobserver variability in ultrasound imaging.

Deep learning is a powerful tool that became practical in 2008, harnessing the power of Graphic Processing Unites, and has developed rapidly in image, video, and natural language processing. There are ongoing developments in the application of deep learning to medical data for a variety of tasks across multiple imaging modalities. The reliability and repeatability of deep learning techniques are of utmost importance if deep learning can be considered a tool for assisting experts, including physicians, radiologists, and sonographers. Owing to the high costs of labeling data, deep learning models are often evaluated against one expert, and it is unknown if any errors fall within a clinically acceptable range. Ultrasound is a commonly used imaging modality for breast cancer screening processes and for visually estimating risk using the Breast Imaging Reporting and Data System score. This process is highly dependent on the skills and experience of the sonographers and radiologists, thereby leading to interobserver variability and interpretation. For these reasons, we propose an interobserver reliability study comparing the performance of a current top-performing deep learning segmentation model against three experts who manually segmented suspicious breast lesions in clinical ultrasound (US) images. We pretrained the model using a US thyroid segmentation dataset with 455 patients and 50,993 images, and trained the model using a US breast segmentation dataset with 733 patients and 29,884 images. We found a mean Fleiss kappa value of 0.78 for the performance of three experts in breast mass segmentation compared to a mean Fleiss kappa value of 0.79 for the performance of experts and the optimized deep learning model.

Automatic Deep Learning Semantic Segmentation of Ultrasound Thyroid Cineclips Using Recurrent Fully Convolutional Networks.

Medical segmentation is an important but challenging task with applications in standardized report generation, remote medicine and reducing medical exam costs by assisting experts. In this paper, we exploit time sequence information using a novel spatio-temporal recurrent deep learning network to automatically segment the thyroid gland in ultrasound cineclips. We train a DeepLabv3+ based convolutional LSTM model in four stages to perform semantic segmentation by exploiting spatial context from ultrasound cineclips. The backbone DeepLabv3+ model is replicated six times and the output layers are replaced with convolutional LSTM layers in an atrous spatial pyramid pooling configuration. Our proposed model achieves mean intersection over union scores of 0.427 for cysts, 0.533 for nodules and 0.739 for thyroid. We demonstrate the potential application of convolutional LSTM models for thyroid ultrasound segmentation.

Automated Segmentation of Thyroid Nodule, Gland, and Cystic Components From Ultrasound Images Using Deep Learning.

Sonographic features associated with margins, shape, size, and volume of thyroid nodules are used to assess their risk of malignancy. Automatically segmenting nodules from normal thyroid gland would enable an automated estimation of these features. A novel multi-output convolutional neural network algorithm with dilated convolutional layers is presented to segment thyroid nodules, cystic components inside the nodules, and normal thyroid gland from clinical ultrasound B-mode scans. A prospective study was conducted, collecting data from 234 patients undergoing a thyroid ultrasound exam before biopsy. The training and validation sets encompassed 188 patients total; the testing set consisted of 48 patients. The algorithm effectively segmented thyroid anatomy into nodules, normal gland, and cystic components. The algorithm achieved a mean Dice coefficient of 0.76, a mean true positive fraction of 0.90, and a mean false positive fraction of 1.61×10-6. The values are on par with a conventional seeded algorithm. The proposed algorithm eliminates the need for a seed in the segmentation process, thus automatically detecting and segmenting the thyroid nodules and cystic components. The detection rate for thyroid nodules and cystic components was 82% and 44%, respectively. The inference time per image, per fold was 107ms. The mean error in volume estimation of thyroid nodules for five select cases was 7.47%. The algorithm can be used for detection, segmentation, size estimation, volume estimation, and generating thyroid maps for thyroid nodules. The algorithm has applications in point of care, mobile health monitoring, improving workflow, reducing localization time, and assisting sonographers with limited expertise.

Multi-parameter Sub-Hertz Analysis of Viscoelasticity With a Quality Metric for Differentiation of Breast Masses.

We applied sub-Hertz analysis of viscoelasticity (SAVE) to differentiate breast masses in pre-biopsy patients. Tissue response during external ramp-and-hold stress was ultrasonically detected. Displacements were used to acquire tissue viscoelastic parameters. The fast instantaneous response and slow creep-like deformations were modeled as the response of a linear standard solid from which viscoelastic parameters were estimated. These parameters were used in a multi-variable classification framework to differentiate malignant from benign masses identified by pathology. When employing all viscoelasticity parameters, SAVE resulted in 71.43% accuracy in differentiating lesions. When combined with ultrasound features and lesion size, accuracy was 82.24%. Adding a quality metric based on uniaxial motion increased the accuracy to 81.25%. When all three were combined with SAVE, accuracy was 91.3%. These results confirm the utility of SAVE as a robust ultrasound-based diagnostic tool for non-invasive differentiation of breast masses when used as stand-alone biomarkers or in conjunction with ultrasonic features.

Predictive value of comb-push ultrasound shear elastography for the differentiation of reactive and metastatic axillary lymph nodes: A preliminary investigation.

OBJECTIVES: To evaluate the predictive performance of comb-push ultrasound shear elastography for the differentiation of reactive and metastatic axillary lymph nodes. METHODS: From June 2014 through September 2018, 114 female volunteers (mean age 58.1±13.3 years; range 28-88 years) with enlarged axillary lymph nodes identified by palpation or clinical imaging were prospectively enrolled in the study. Mean, standard deviation and maximum shear wave elastography parameters from 117 lymph nodes were obtained and compared to fine needle aspiration biopsy results. Mann-Whitney U test and ROC curve analysis were performed. RESULTS: The axillary lymph nodes were classified as reactive or metastatic based on the fine needle aspiration outcomes. A statistically significant difference between reactive and metastatic axillary lymph nodes was observed based on comb-push ultrasound shear elastography (CUSE) results (p<0.0001) from mean and maximum elasticity values. Mean elasticity showed the best separation with a ROC analysis resulting in 90.5% sensitivity, 94.4% specificity, 0.97 area under the curve, 95% positive predictive value, and 89.5% negative predictive value with a 30.2-kPa threshold. CONCLUSIONS: CUSE provided a quantifiable parameter that can be used for the assessment of enlarged axillary lymph nodes to differentiate between reactive and metastatic processes.

Ultrasensitive Ultrasound Microvessel Imaging for Characterizing Benign and Malignant Breast Tumors.

Tumor angiogenesis plays an important role during breast tumor growth. However, conventional Doppler has limited sensitivity to detect small blood vessels, resulting in a large overlap of Doppler features between benign and malignant tumors. An ultrasensitive ultrasound microvessel imaging (UMI) technique was recently developed. To evaluate the performance of UMI, we studied 44 patients with 51 breast masses. Tumor pathology served as the gold standard: 28 malignancies and 23 benignities. UMI provided a significant improvement in depicting smaller vessels compared with conventional Doppler. The microvessel morphologies observed on UMI were associated with tumor benign/malignant classification. The diagnostic accuracy of correct Breast Imaging Reporting and Data System (BI-RADS) classification rate (BI-RADS ≥4a: test positive; BI-RADS ≤3: test negative) as a fraction of total mass population was improved by 16% after combining conventional ultrasound with UMI compared with using conventional ultrasound alone. This improvement indicates the potential of UMI in reducing unnecessary benign biopsies and avoiding missed malignant biopsies.

Automated and real-time segmentation of suspicious breast masses using convolutional neural network.

In this work, a computer-aided tool for detection was developed to segment breast masses from clinical ultrasound (US) scans. The underlying Multi U-net algorithm is based on convolutional neural networks. Under the Mayo Clinic Institutional Review Board protocol, a prospective study of the automatic segmentation of suspicious breast masses was performed. The cohort consisted of 258 female patients who were clinically identified with suspicious breast masses and underwent clinical US scan and breast biopsy. The computer-aided detection tool effectively segmented the breast masses, achieving a mean Dice coefficient of 0.82, a true positive fraction (TPF) of 0.84, and a false positive fraction (FPF) of 0.01. By avoiding positioning of an initial seed, the algorithm is able to segment images in real time (13-55 ms per image), and can have potential clinical applications. The algorithm is at par with a conventional seeded algorithm, which had a mean Dice coefficient of 0.84 and performs significantly better (P< 0.0001) than the original U-net algorithm.

Differentiation of Benign and Malignant Thyroid Nodules by Using Comb-push Ultrasound Shear Elastography: A Preliminary Two-plane View Study.

RATIONALE AND OBJECTIVES: Low specificity of traditional ultrasound in differentiating benign from malignant thyroid nodules leads to a great number of unnecessary (ie, benign) fine-needle aspiration biopsies that causes a significant financial and physical burden to the patients. Ultrasound shear wave elastography is a technology capable of providing additional information related to the stiffness of tissues. In this study, quantitative stiffness values acquired by ultrasound shear wave elastography in two different imaging planes were evaluated for the prediction of malignancy in thyroid nodules. In addition, the association of elasticity measurements with sonographic characteristics of thyroid gland and nodules is presented. MATERIALS AND METHODS: A total number of 155 patients (106 female and 49 male) (average age 57.48 ± 14.44 years) with 173 thyroid nodules (average size 24.89 ± 15.41 mm, range 5-68 mm) scheduled for fine-needle aspiration biopsy were recruited from March 2015 to May 2017. Comb-push shear elastography imaging was performed at longitudinal and transverse anatomic planes. Mean (Emean) and maximum (Emax) elasticity values were obtained. RESULTS: Measurements at longitudinal view were statistically significantly higher than measurements at transverse view. Nodules with calcifications were associated with increased elasticity, and nodules with a vascular component or within an enlarged thyroid gland (goiter) were associated with a lower elasticity value. Receiver operating characteristic curve analysis was performed for Emean and Emax at each imaging plane and for the average of both planes. Sensitivity of 95.45%, specificity of 86.61%, 0.58 positive predictive value, and 0.99 negative predictive value were achieved by the average of the two planes for each Emean and Emax parameters, with area under the curve of 92% and 93%, and a cutoff value of 49.09 kPa and 105.61 kPa, respectively. CONCLUSIONS: The elastic properties of thyroid nodules showed promise to be a good discriminator between malignant and benign nodules (P < .0001). However, probe orientation and internal features such as calcifications, vascular component, and goiter may influence the final elastography measurements. A larger number of malignant nodules need to be studied to further validate our results.

Comb-push ultrasound shear elastography of breast masses: initial results show promise.

PURPOSE OR OBJECTIVE: To evaluate the performance of Comb-push Ultrasound Shear Elastography (CUSE) for classification of breast masses. MATERIALS AND METHODS: CUSE is an ultrasound-based quantitative two-dimensional shear wave elasticity imaging technique, which utilizes multiple laterally distributed acoustic radiation force (ARF) beams to simultaneously excite the tissue and induce shear waves. Female patients who were categorized as having suspicious breast masses underwent CUSE evaluations prior to biopsy. An elasticity estimate within the breast mass was obtained from the CUSE shear wave speed map. Elasticity estimates of various types of benign and malignant masses were compared with biopsy results. RESULTS: Fifty-four female patients with suspicious breast masses from our ongoing study are presented. Our cohort included 31 malignant and 23 benign breast masses. Our results indicate that the mean shear wave speed was significantly higher in malignant masses (6 ± 1.58 m/s) in comparison to benign masses (3.65 ± 1.36 m/s). Therefore, the stiffness of the mass quantified by the Young's modulus is significantly higher in malignant masses. According to the receiver operating characteristic curve (ROC), the optimal cut-off value of 83 kPa yields 87.10% sensitivity, 82.61% specificity, and 0.88 for the area under the curve (AUC). CONCLUSION: CUSE has the potential for clinical utility as a quantitative diagnostic imaging tool adjunct to B-mode ultrasound for differentiation of malignant and benign breast masses.

Comb-push ultrasound shear elastography (CUSE) for evaluation of thyroid nodules: preliminary in vivo results.

In clinical practice, an overwhelming majority of biopsied thyroid nodules are benign. Therefore, there is a need for a complementary and noninvasive imaging tool to provide clinically relevant diagnostic information about thyroid nodules to reduce the rate of unnecessary biopsies. The goal of this study was to evaluate the feasibility of utilizing comb-push ultrasound shear elastography (CUSE) to measure the mechanical properties (i.e., stiffness) of thyroid nodules and use this information to help classify nodules as benign or malignant. CUSE is a fast and robust 2-D shear elastography technique in which multiple laterally distributed acoustic radiation force beams are utilized simultaneously to produce shear waves. Unlike other shear elasticity imaging modalities, CUSE does not suffer from limited field of view (FOV) due to shear wave attenuation and can provide a large FOV at high frame rates. To evaluate the utility of CUSE in thyroid imaging, a preliminary study was performed on a group of five healthy volunteers and 10 patients with ultrasound-detected thyroid nodules prior to fine needle aspiration biopsy. The measured shear wave speeds in normal thyroid tissue and thyroid nodules were converted to Young's modulus (E), indicating a measure of tissue stiffness. Our results indicate an increase in E for thyroid nodules compared to normal thyroid tissue. This increase was significantly higher in malignant nodules compared to benign. The Young's modulus in normal thyroid tissue, benign and malignant nodules were found to be 23.2 ±8.29 kPa, 91.2±34.8 kPa, and 173.0±17.1 kPa, respectively. Results of this study suggest the utility of CUSE in differentiating between benign and malignant thyroid nodules.

Noninvasive assessment of liver fibrosis using ultrasound-based shear wave measurement and comparison to magnetic resonance elastography.

OBJECTIVES: Magnetic resonance elastography (MRE) has excellent performance in detecting liver fibrosis and is becoming an alternative to liver biopsy in clinical practice. Ultrasound techniques based on measuring the propagation speed of the shear waves induced by acoustic radiation force also have shown promising results for liver fibrosis staging. The objective of this study was to compare ultrasound-based shear wave measurement to MRE. METHODS: In this study, 50 patients (28 female and 22 male; age range, 19-81 years) undergoing liver MRE examinations were studied with an ultrasound scanner modified with shear wave measurement functionality. For each patient, 27 shear wave speed measurements were obtained at various locations in the liver parenchyma away from major vessels. The median shear wave speed from all measurements was used to calculate a representative shear modulus (µ) for each patient. Magnetic resonance elastographic data processing was done by a single analyst blinded to the ultrasound measurement results. RESULTS: Ultrasound and MRE measurements were correlated (r = 0.86; P < .001). Receiver operating characteristic (ROC) analysis was applied to the ultrasound measurement results with the MRE diagnosis as the "ground truth." The area under the ROC curve for separating patients with minimum fibrosis (defined as µ(MRE) ≤2.9 kPa) was 0.89 (95% confidence interval, 0.77-0.95), and the area under the ROC curve for separating patients with advanced fibrosis (defined as µ(MRE) ≥5.0 kPa) was 0.96 (95% confidence interval, 0.87-0.99). CONCLUSIONS: Results indicate that the ultrasound-based shear wave measurement correlates with MRE and is a promising method for liver fibrosis staging.

External vibration multi-directional ultrasound shearwave elastography (EVMUSE): application in liver fibrosis staging.

Shear wave speed can be used to assess tissue elasticity, which is associated with tissue health. Ultrasound shear wave elastography techniques based on measuring the propagation speed of the shear waves induced by acoustic radiation force are becoming promising alternatives to biopsy in liver fibrosis staging. However, shear waves generated by such methods are typically very weak. Therefore, the penetration may become problematic, especially for overweight or obese patients. In this study, we developed a new method called external vibration multi-directional ultrasound shearwave elastography (EVMUSE), in which external vibration from a loudspeaker was used to generate a multi-directional shear wave field. A directional filter was then applied to separate the complex shear wave field into several shear wave fields propagating in different directions. A 2-D shear wave speed map was reconstructed from each individual shear wave field, and a final 2-D shear wave speed map was constructed by compounding these individual wave speed maps. The method was validated using two homogeneous phantoms and one multi-purpose tissue-mimicking phantom. Ten patients undergoing liver magnetic resonance elastography (MRE) were also studied with EVMUSE to compare results between the two methods. Phantom results showed EVMUSE was able to quantify tissue elasticity accurately with good penetration. In vivo EVMUSE results were well correlated with MRE results, indicating the promise of using EVMUSE for liver fibrosis staging.

Breast lesions: evaluation with US strain imaging--clinical experience of multiple observers.

PURPOSE: To prospectively determine the accuracy of using an ultrasonographic (US) strain imaging technique known as lesion size comparison to differentiate benign from malignant breast lesions. MATERIALS AND METHODS: Institutional Review Board approval and patient informed consent were obtained for this HIPPA-compliant study. US strain imaging was performed prospectively for 89 breast lesions in 88 patients. Lesions were imaged by using freehand compression and a real-time strain imaging algorithm. Five observers obtained manual measurements of lesion height, width, and area from B-mode and strain images. By using these size measurements, individual observer and group performances were assessed by using the area under the receiver operating characteristic curve (A(z)). The performance of a single size parameter versus that of a combination of size parameters was evaluated by using univariate and multivariate logistic regression. RESULTS: Group A(z) values showed that width ratio and area ratio yielded the best results for differentiating benign and malignant breast lesions, and they were not statistically different from one another (P = .499). For the group, the performance of area and width, which was superior to that of height and aspect ratio, was statistically significant for all cases (P < .011) except for those that compared area with aspect ratio (P = .118). By using a group threshold of 1.04 for width ratio and 1.13 for area ratio, the sensitivity and specificity of the technique were 96% and 21%, respectively, for width and 96% and 24%, respectively, for area. The best observer achieved a sensitivity of 96% and a specificity of 61% by using the area ratio. For all but one observer, combined size parameters did not improve observer performance (P > .258). Significant interobserver performance variability was observed (P < .001). CONCLUSION: Results suggest that US strain imaging has the potential to aid diagnosis of breast lesions. However, manually tracing lesion boundaries for size ratio differentiation in a busy clinical setting did not match the diagnostic performance levels previously reported. Focusing on measurements of lesion width, along with additional observer training or automated processes, may yield a suitable method for routine clinical application.