Theodore W. Cary

Teaching Medical Students Diagnostic Sonography

The purpose of this pilot project was to train medical students in sonography.

Computer-based margin analysis of breast sonography for differentiating malignant and benign masses.

Objective. To evaluate the role of quantitative margin features in the computer‐aided diagnosis of malignant and benign solid breast masses using sonographic imaging. Methods. Sonographic images from 56 patients with 58 biopsy‐proven masses were analyzed quantitatively for the following features: margin sharpness, margin echogenicity, and angular variation in margin. Of the 58 masses, 38 were benign and 20 were malignant. Each feature was evaluated individually and in combination with the others to determine its association with malignancy. The combination of features yielding the highest association with malignancy was analyzed by logistic regression to determine the probability of malignancy. The performance of the probability measurements was evaluated by receiver operating characteristic analysis using a round‐robin technique. Results. Margin sharpness, margin echogenicity, and angular variation in margin were significantly different for the malignant and benign masses (P < .03, 2‐tailed Student t test). According to quantitative measures, tumor‐tissue margins of the malignant masses were less distinct than for the benign masses. Although the mean size of the lesions for the two groups was the same, the mean age of the patients was statistically different (P = .000625). After logistic regression analysis, the individual features age, margin sharpness, margin echogenicity, and angular variation in margin were found to be associated with the probability of malignancy (P < .03). The area under the receiver operating characteristic curve ± SD for the 3‐feature logistic regression model combining age, margin echogenicity, and angular variation of margin was 0.87 ± 0.05. Conclusions. The proposed quantitative margin features are robust and can reliably measure margin distinctiveness. These features combined with logistic regression analysis can be useful for computer‐aided diagnosis of solid breast lesions.

In Vivo Noninvasive Temperature Measurement by B-Mode Ultrasound Imaging

Objective. This study investigated the use of ultrasound image analysis in quantifying temperature changes in tissue, both ex vivo and in vivo, undergoing local hyperthermia. Methods. Temperature estimation is based on the thermal dependence of the acoustic speed in a heated medium. Because standard beam‐forming algorithms on clinical ultrasound scanners assume a constant acoustic speed, temperature‐induced changes in acoustic speed produce apparent scatterer displacements in B‐mode images. A cross‐correlation algorithm computes axial speckle pattern displacement in B‐mode images of heated tissue, and a theoretically derived temperature‐displacement relationship is used to generate maps of temperature changes within the tissue. Validation experiments were performed on excised tissue and in murine subjects, wherein low‐intensity ultrasound was used to thermally treat tissue for several minutes. Diagnostic temperature estimation was performed using a linear array ultrasound transducer, while a fine‐wire thermocouple invasively measured the temperature change. Results. Pearson correlations ± SDs between the image‐derived and thermocouple‐measured temperature changes were R2 = 0.923 ± 0.066 for 4 thermal treatments of excised bovine muscle tissue and R2 = 0.917 ± 0.036 for 4 treatments of in vivo murine tumor tissue. The average differences between the two temperature measurements were 0.87°C ± 0.72°C for ex vivo studies and 0.97°C ± 0.55°C for in vivo studies. Maps of the temperature change distribution in tissue were generated for each experiment. Conclusions. This study demonstrates that velocimetric measurement on B‐mode images has potential to assess temperature changes noninvasively in clinical applications.

Multigated contrast-enhanced power Doppler to measure blood flow in mice tumors

This study evaluated an image-gating method using contrast-enhanced power Doppler ultrasound (US) to estimate blood perfusion in mice tumors. A mathematical model that compensates for the effect of bubble destruction by US pulses was used to determine contrast flow through an image plane. Multigated power Doppler images were obtained following contrast injection. Contrast flow index (CFI) was determined by measuring the area under the color level vs. time curve for each gating frequency. CFI was compared with true flow. The method was first evaluated using a flow phantom with variable flow rates, and then verified in a mouse model with implanted tumors. Color levels in Doppler images were modulated with gating frequency due to variable destruction of microbubbles by US pulses. CFI measured from the images correlated strongly with true flow in the flow phantom (r(2) = 0.87). The proposed method yielded reproducible CFI for mice tumors, suggesting that multigated contrast-enhanced power Doppler imaging may provide noninvasive measurement of tumor perfusion in mice.

Delta-projection imaging on contrast-enhanced ultrasound to quantify tumor microvasculature and perfusion.

RATIONALE AND OBJECTIVES The aim of this study was to assess the Delta-projection image processing technique for visualizing tumor microvessels and for quantifying the area of tissue perfused by them on contrast-enhanced ultrasound images. MATERIALS AND METHODS The Delta-projection algorithm was implemented to quantify perfusion by tracking the running maximum of the difference (Delta) between the contrast-enhanced ultrasound image sequence and a baseline image. Twenty-five mice with subcutaneous K1735 melanomas were first imaged with contrast-enhanced grayscale and then with minimum-exposure contrast-enhanced power Doppler (minexCPD) ultrasound. Delta-projection images were reconstructed from the grayscale images and then used to evaluate the evolution of tumor vascularity during the course of contrast enhancement. The extent of vascularity (ratio of the perfused area to the tumor area) for each tumor was determined quantitatively from Delta-projection images and compared to the extent of vascularity determined from contrast-enhanced power Doppler images. Delta-projection and minexCPD measurements were compared using linear regression analysis. RESULTS Delta-projection was successfully performed in all 25 cases. The technique allowed the dynamic visualization of individual blood vessels as they filled in real time. Individual tumor blood vessels were distinctly visible during early image enhancement. Later, as an increasing number of blood vessels were filled with the contrast agent, clusters of vessels appeared as regions of perfusion, and the identification of individual vessels became difficult. Comparisons were made between the perfused area of tumors in Delta-projections and in minexCPD images. The Delta-projection perfusion measurements were correlated linearly with minexCPD. CONCLUSION Delta-projection visualized tumor vessels and enabled the quantitative assessment of the tumor area perfused by the contrast agent.

Brachial artery vasomotion and transducer pressure effect on measurements by active contour segmentation on ultrasound

PURPOSE To use feed-forward active contours (snakes) to track and measure brachial artery vasomotion on ultrasound images recorded in both transverse and longitudinal views; and to compare the algorithms performance in each view. METHODS Longitudinal and transverse view ultrasound image sequences of 45 brachial arteries were segmented by feed-forward active contour (FFAC). The segmented regions were used to measure vasomotion artery diameter, cross-sectional area, and distention both as peak-to-peak diameter and as area. ECG waveforms were also simultaneously extracted frame-by-frame by thresholding a running finite-difference image between consecutive images. The arterial and ECG waveforms were compared as they traced each phase of the cardiac cycle. RESULTS FFAC successfully segmented arteries in longitudinal and transverse views in all 45 cases. The automated analysis took significantly less time than manual tracing, but produced superior, well-behaved arterial waveforms. Automated arterial measurements also had lower interobserver variability as measured by correlation, difference in mean values, and coefficient of variation. Although FFAC successfully segmented both the longitudinal and transverse images, transverse measurements were less variable. The cross-sectional area computed from the longitudinal images was 27% lower than the area measured from transverse images, possibly due to the compression of the artery along the image depth by transducer pressure. CONCLUSIONS FFAC is a robust and sensitive vasomotion segmentation algorithm in both transverse and longitudinal views. Transverse imaging may offer advantages over longitudinal imaging: transverse measurements are more consistent, possibly because the method is less sensitive to variations in transducer pressure during imaging.

Comparison of naive Bayes and logistic regression for computer-aided diagnosis of breast masses using ultrasound imaging

This study compares the performance of two proven but very different machine learners, Naïve Bayes and logistic regression, for differentiating malignant and benign breast masses using ultrasound imaging. Ultrasound images of 266 masses were analyzed quantitatively for shape, echogenicity, margin characteristics, and texture features. These features along with patient age, race, and mammographic BI-RADS category were used to train Naïve Bayes and logistic regression classifiers to diagnose lesions as malignant or benign. ROC analysis was performed using all of the features and using only a subset that maximized information gain. Performance was determined by the area under the ROC curve, Az, obtained from leave-one-out cross validation. Naïve Bayes showed significant variation (Az 0.733 ± 0.035 to 0.840 ± 0.029, P < 0.002) with the choice of features, but the performance of logistic regression was relatively unchanged under feature selection (Az 0.839 ± 0.029 to 0.859 ± 0.028, P = 0.605). Out of 34 features, a subset of 6 gave the highest information gain: brightness difference, margin sharpness, depth-to-width, mammographic BI-RADs, age, and race. The probabilities of malignancy determined by Naïve Bayes and logistic regression after feature selection showed significant correlation (R2= 0.87, P < 0.0001). The diagnostic performance of Naïve Bayes and logistic regression can be comparable, but logistic regression is more robust. Since probability of malignancy cannot be measured directly, high correlation between the probabilities derived from two basic but dissimilar models increases confidence in the predictive power of machine learning models for characterizing solid breast masses on ultrasound.

Combined Naïve Bayes and logistic regression for quantitative breast sonography

Sonography is commonly used as an adjunct to mammography for early detection of breast cancer. We are developing methods to classify solid breast masses in sonograms as malignant or benign. The goal of this study was to combine two independent probabilistic classifiers to improve computer-aided diagnosis of breast masses. Naïve Bayes and logistic regression were used for supervised classification of masses from extracted morphological sonographic features, in combination with mammographic BI-RADS (categories 1 to 5) and patient age. Solid masses with biopsy-proven diagnoses were analyzed. Training and testing were performed using leave-one-out cross validation. Diagnostic performance was evaluated by the area under the curve (AUC) of the receiver operating characteristic (ROC). Agreement between predictions from the two classifiers was used to differentiate benign and malignant masses. The results show that logistic regression and Naïve Bayes performed with ROC area of 0.902 ± 0.023 and 0.865 ± 0.027, respectively. The combined use of logistic regression and Naïve Bayes demonstrated reduction in biopsies by 48%, with malignancy missed in 2% of cases (false negative rate of 6.4%).

The diagnostic performance of leak-plugging automated segmentation versus manual tracing of breast lesions on ultrasound images:

Purpose To assess the diagnostic performance of a leak-plugging segmentation method that we have developed for delineating breast masses on ultrasound images. Materials and methods Fifty-two biopsy-proven breast lesion images were analyzed by three observers using the leak-plugging and manual segmentation methods. From each segmentation method, grayscale and morphological features were extracted and classified as malignant or benign by logistic regression analysis. The performance of leak-plugging and manual segmentations was compared by: size of the lesion, overlap area (Oa) between the margins, and area under the ROC curves (Az). Results The lesion size from leak-plugging segmentation correlated closely with that from manual tracing (R2 of 0.91). Oa was higher for leak plugging, 0.92 ± 0.01 and 0.86 ± 0.06 for benign and malignant masses, respectively, compared to 0.80 ± 0.04 and 0.73 ± 0.02 for manual tracings. Overall Oa between leak-plugging and manual segmentations was 0.79 ± 0.14 for benign and 0.73 ± 0.14 for malignant lesions. Az for leak plugging was consistently higher (0.910 ± 0.003) compared to 0.888 ± 0.012 for manual tracings. The coefficient of variation of Az between three observers was 0.29% for leak plugging compared to 1.3% for manual tracings. Conclusion The diagnostic performance, size measurements, and observer variability for automated leak-plugging segmentations were either comparable to or better than those of manual tracings.

Machine learning for diagnostic ultrasound of triple-negative breast cancer

PurposeEarly diagnosis of triple-negative (TN) breast cancer is important due to its aggressive biological characteristics, poor clinical outcomes, and limited options for therapy. The goal of this study is to evaluate the potential of machine learning with quantitative ultrasound image features for the diagnosis of TN breast cancer.MethodsUltrasonic and clinical data of 140 surgically confirmed breast cancer cases were analyzed retrospectively for the diagnosis of TN and non-TN (NTN) subtypes. The subtypes were classified based on the expression of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2). Ultrasound image features were measured from the grayscale and color Doppler images and used with logistic regression for classification by machine learning. Leave-one-out cross validation was used to train and test the differentiation. Diagnostic performance was measured by the area under receiver operating characteristic (ROC) curve, and sensitivity and specificity determined at the Youdons index.ResultsOf the twelve grayscale and Doppler features measured, eight were found to be statistically different for the TN and NTN subtypes (p < 0.05). The area under the ROC curve (AUC) of the statistically significant grayscale (GS) and color Doppler (CD) features was 0.85 and 0.65, respectively. The AUC increased to 0.88 when the GS and CD features were used together, with sensitivity of 86.96% and specificity of 82.91%. Consideration of patient age in the analysis did not improve discrimination of TN and NTN.ConclusionsThe analysis of breast ultrasound images by machine learning achieves high level of differentiation between the TN and NTN subtypes, exceeding the diagnostic performance by standard visual assessments of the images.