To evaluate the efficacy of contrast-enhanced ultrasound (CEUS) lymphosonography in the identification of sentinel lymph nodes (SLNs) in patients with breast cancer undergoing surgical excision.*Methods and Materials To date 66 women scheduled for breast cancer surgery with SLN excision were enrolled in this ongoing, IRB-approved study. Subjects underwent percutaneous Sonazoid (GE Healthcare, Oslo, Norway) injections around the tumor area at the 12,-3,-6,-and 9 o’clock positions for a total of 1.0 ml. Lymphosonography was done using Cadence Pulse Sequencing on an S3000 HELX scanner (Siemens Healthineers, Mountain View, CA) with a 9L4 linear probe. Subjects received blue dye and radioactive tracer as part of their standard of care. After surgical excision, the ex-vivo SLNs specimen were scanned using color Doppler to confirm the uptake of Sonazoid and sent for pathology. The excised SLNs were classified as positive or negative for presence of blue dye, radioactive tracer and Sonazoid. The results were compared between methods and pathology findings.*Results One-hundred and ninety-five SLNs were surgically excised from 66 subjects, 125 were positive for blue dye, 175 were positive for the radioactive tracer and 166 were positive for Sonazoid. Comparison with the reference standard (blue dye) showed that the radioactive tracer had an accuracy of 67%, while lymphosonography achieved an accuracy of 73% (p=0.56). When the comparison was done with radioactive tracer as the reference standard, the blue dye injections had an accuracy of 66%, while lymphosonography achieved an accuracy of 80% (p<0.001). Of the 195 SLNs excised, 26 were determined to be malignant by pathology; amongst them 13 were positive for blue dye, 19 were positive for radioactive tracer and 26 were positive for Sonazoid, which translated into an accuracy of 50% for blue dye, 73% for radioactive tracer and 100% for lymphosonography (p<0.008).*Conclusions Lymphosonography achieved better accuracy compared with radioactive tracer and blue dye for identifying SLNs in breast cancer patients. All the 26 SLNs positive for malignancy were identified by lymphosonography.*Clinical Relevance/Application Lymphosonography is an ultrasound modality that uses ultrasound contrast agents to identify SLNs, which is an important aspect of predicting outcomes for patients with breast cancer.

One-hundred and ninety-five SLNs were surgically excised from 66 subjects, 125 were positive for blue dye, 175 were positive for the radioactive tracer and 166 were positive for Sonazoid. Comparison with the reference standard (blue dye) showed that the radioactive tracer had an accuracy of 67%, while lymphosonography achieved an accuracy of 73% (p=0.56). When the comparison was done with radioactive tracer as the reference standard, the blue dye injections had an accuracy of 66%, while lymphosonography achieved an accuracy of 80% (p<0.001). Of the 195 SLNs excised, 26 were determined to be malignant by pathology; amongst them 13 were positive for blue dye, 19 were positive for radioactive tracer and 26 were positive for Sonazoid, which translated into an accuracy of 50% for blue dye, 73% for radioactive tracer and 100% for lymphosonography (p<0.008).

Lymphosonography is an ultrasound modality that uses ultrasound contrast agents to identify SLNs, which is an important aspect of predicting outcomes for patients with breast cancer.

To assess the efficacy of percutaneous Sonazoid-enhanced ultrasound and in vitro verification for identification sentinel lymph nodes (SLNs) and diagnosis of metastatic SLNs in patients with early breast cancer(BC).*Methods and Materials 115 patients with early BC were enrolled finally. After the induction of general anesthesia, 0.4 ml of Sonazoid (SNZ), a new second-generation tissue specific ultrasound contrast agent (UCA), mixed with 0.6ml of methylene blue, was injected intradermally. The lymphatic vessels and connected SLNs were immediately observed and marked. After being resected, these SLNs were soaked in saline water and examined still in the mode of contrast enhanced ultrasound (CEUS) in vitro, This procedure could ensure that all the enhanced nodes had been removed as much as possible. The numbers of SLNs detected by UCA and blue dye were recorded. The enhancement patterns of SLNs were compared with the final pathological results.*Results SLNs detection rate by SNZ-CEUS was 94.78%. CEUS identified a median of 1.5 nodes, while blue dye identified a median of 1.84 nodes per case(P<0.001). When homogeneous high perfusion, complete annular high perfusion and low perfusion were considered negative, the diagnostic sensitivity and specificity of CEUS for SLN were 90.91% and 84.50%, and the negative predictive value was 97.32%. The FNR was 9.09%. The diagnostic performance of CEUS for SLN is better than that of gray scale ultrasonography.*Conclusions Percutaneous SNZ-enhanced ultrasonography combined with in vitro verification is a feasible and reliable method for SLNs identification intraoperatively. Enhancement patterns can be helpful in determining the status of SLNs.*Clinical Relevance/Application (1)CEUS with percutaneous injection of Sonazoid can successfully and accurately identify and character SLNs in early breast cancer patients. (2) Sonazoid, with high affinity with reticuloendothelial cells, increases the imaging time of SLNs and facilitates biopsy intraoperatively better than Sonovue as a lymphatic tracer.

SLNs detection rate by SNZ-CEUS was 94.78%. CEUS identified a median of 1.5 nodes, while blue dye identified a median of 1.84 nodes per case(P<0.001). When homogeneous high perfusion, complete annular high perfusion and low perfusion were considered negative, the diagnostic sensitivity and specificity of CEUS for SLN were 90.91% and 84.50%, and the negative predictive value was 97.32%. The FNR was 9.09%. The diagnostic performance of CEUS for SLN is better than that of gray scale ultrasonography.

(1)CEUS with percutaneous injection of Sonazoid can successfully and accurately identify and character SLNs in early breast cancer patients. (2) Sonazoid, with high affinity with reticuloendothelial cells, increases the imaging time of SLNs and facilitates biopsy intraoperatively better than Sonovue as a lymphatic tracer.

Breast cancer is the second most common cancer in the world and the most frequent type of cancer among women (30% of all cancers). This multi-center study assessed the ability of contrast-enhanced, nonlinear 3D US imaging to characterize previously indeterminate breast lesions using quantitative parameters and clinical assessments.*Methods and Materials In total 236 women with biopsy-proven breast lesions were enrolled in this prospective, FDA approved study (IND: 112,241). Following conventional US and power Doppler imaging (PDI), an US contrast agent (Definity, Lantheus Medical Imaging, N Billerica, MA) was administrated IV. Contrast-enhanced 3D harmonic imaging (HI; transmitting/receiving at 5.0/10.0 MHz) as well as 3D subharmonic imaging (SHI; transmitting/receiving at 5.8/2.9 MHz) were performed using a modified Logiq 9 scanner (GE Healthcare, Waukesha, WI) with a 4D10L probe. Five radiologists blinded to other results independently scored the 4 randomized US modes using a 7-point BIRADS scale from negative to highly suggestive of malignancy as well as lesion vascularity and diagnostic confidence. Quantitative parametric volumes were constructed from time-intensity curves for vascular heterogeneity, perfusion and area under the curve. Diagnostic accuracy for US and mammography were determined relative to pathology using ROC and reverse, step-wise logistical regression analyses. The ? statistic was calculated for inter-reader agreement.*Results Of the 236 cases, 219 were successfully scanned and biopsies indicated 164 (75%) benign and 55 (25%) malignant lesions. 3D HI showed flow in 8 lesions, whereas 3D SHI visualized flow in 83 lesions. SHI depicted more anastomoses and vascularity than HI (p<0.021), but there were no differences by pathology (p>0.27). US modes achieved accuracies from 79-85%, which was significantly better than mammography (72%; p<0.03). SHI increased diagnostic confidence by 3-6% (p<0.01), but inter-reader agreements were medium to low (?<0.52). The best logistical regression model achieved a 96% accuracy by combining clinical reads and quantitative 3D SHI parameters.*Conclusions 3D SHI is better at detecting contrast flow in vascular breast masses than 3D HI. Characterization of indeterminate breast lesions with quantitative 3D SHI parameters and clinical assessments improves diagnostic accuracy.*Clinical Relevance/Application Combining quantitative 3D SHI parameters and radiologists’ assessments increase the accuracy and confidence for characterizing indeterminate breast lesions.

Of the 236 cases, 219 were successfully scanned and biopsies indicated 164 (75%) benign and 55 (25%) malignant lesions. 3D HI showed flow in 8 lesions, whereas 3D SHI visualized flow in 83 lesions. SHI depicted more anastomoses and vascularity than HI (p<0.021), but there were no differences by pathology (p>0.27). US modes achieved accuracies from 79-85%, which was significantly better than mammography (72%; p<0.03). SHI increased diagnostic confidence by 3-6% (p<0.01), but inter-reader agreements were medium to low (?<0.52). The best logistical regression model achieved a 96% accuracy by combining clinical reads and quantitative 3D SHI parameters.

Combining quantitative 3D SHI parameters and radiologists’ assessments increase the accuracy and confidence for characterizing indeterminate breast lesions.

To train an AI system to triage breast US exams into an enhanced assessment workflow and a no radiologist workflow (standalone AI interpretation of US exams with very low probability for malignancy), with the goal of reallocating radiologist time towards exams with high suspicion of malignancy.*Methods and Materials Our AI model was based on a neural network inspired by the Globally-Aware Multiple Instance Classifier. To develop and validate this system, we curated a dataset consisting of 288,767 breast US exams with 5,442,907 total images acquired from 143,203 patients examined between 2012 and 2019 at a large academic medical center. 28,914 of these exams were associated with at least one biopsy procedure, 5,593 of which had biopsies yielding malignant findings. Pathology was used as the reference standard. This dataset was split on into training (60%), validation (10%), and test datasets (30%). The AI system was initially trained to automatically detect and classify breast lesions on US imaging with imaging-level data and did not require region of interest input from radiologists. Predictions from this system were then used to channel women to the two new workflows: a no radiologist workflow and an enhanced assessment workflow.*Results On a test set of 44,755 exams, the AI system achieved an AUC of 0.976 for identifying exams with malignant lesions. When triaging 60%, 70%, or 80% of women with the lowest AI scores from the test set into the no radiologist workflow, the false negative rate of the AI system was 1 out every 11905 exams (0.008%), 4608 exams (0.02%), and 2532 exams (0.04%) respectively. When this triage system was used to evaluate 3553 exams from the test set which were given a BI-RADS 3 assessment at the time of original read, it reclassified 60%, 70%, and 80% of exams with the lowest AI scores as benign without missing any malignant lesions. When the AI system utilized a high specificity threshold to triage exams it considered to be at high risk for malignancy, it placed 978 (2.2% of total) exams into an enhanced assessment workflow, with high positive predictive value (69.6%) that significantly exceeded that of the breast radiologists who initially evaluated the test set exams (15.7%). Despite representing only 2.2% of the test dataset, this enhanced assessment workflow contained 56.5% of all malignant cases in the dataset.*Conclusions Our AI system could potentially eliminate up to 60-80% of breast US exams from the radiologist worklist, with a false-negative rate of 0.008-0.04%.*Clinical Relevance/Application Using a high sensitivity threshold, AI based software may function as a standalone interpreter and eliminate low probability of malignancy cases from the radiologist worklist.

On a test set of 44,755 exams, the AI system achieved an AUC of 0.976 for identifying exams with malignant lesions. When triaging 60%, 70%, or 80% of women with the lowest AI scores from the test set into the no radiologist workflow, the false negative rate of the AI system was 1 out every 11905 exams (0.008%), 4608 exams (0.02%), and 2532 exams (0.04%) respectively. When this triage system was used to evaluate 3553 exams from the test set which were given a BI-RADS 3 assessment at the time of original read, it reclassified 60%, 70%, and 80% of exams with the lowest AI scores as benign without missing any malignant lesions. When the AI system utilized a high specificity threshold to triage exams it considered to be at high risk for malignancy, it placed 978 (2.2% of total) exams into an enhanced assessment workflow, with high positive predictive value (69.6%) that significantly exceeded that of the breast radiologists who initially evaluated the test set exams (15.7%). Despite representing only 2.2% of the test dataset, this enhanced assessment workflow contained 56.5% of all malignant cases in the dataset.

Using a high sensitivity threshold, AI based software may function as a standalone interpreter and eliminate low probability of malignancy cases from the radiologist worklist.

To evaluate an AI system to assist radiologists in interpreting breast US exams.*Methods and Materials We developed and evaluated an AI system using a DCNN inspired by the Globally-Aware Multiple Instance Classifier. The model automatically identified malignant and benign lesions without requiring manual annotations from radiologists. The AI system was trained using our internal dataset of 288,767 US exams with 5,442,907 total images acquired from 143,203 patients between 2012-2019, including screening and diagnostic exams. 28,914 of these exams were associated with at least one biopsy procedure, 5,593 of which had biopsies yielding malignant findings. Pathology was used as the reference standard. This dataset was split on a patient level into training (60%), validation (10%), and test datasets (30%). We validated our AI system with a reader study with 10 board-certified breast radiologists (average 14.5 years of experience). Each reader reviewed 663 exams that were sampled from the test set. A hybrid decision-making model was created for each reader which made predictions by evenly weighting the predictions of the reader and AI system. Diagnostic accuracy of the AI system, readers, and hybrid models were evaluated using ROC curves.*Results On a test set of 44,755 exams, the AI system achieved an AUC of 0.976 for identifying exams with malignancies. Among the 663 reader study exams, the AI system had an AUC of 0.962, which was significantly higher than the average radiologist (0.929 ± 0.018, p < 0.001). At the average radiologist’s sensitivity (90.1%), the AI system had a higher specificity (85.6% vs 80.7%, p < 0.001) and recommended fewer biopsies (19.8% vs 24.3%, p < 0.001). On average, the hybrid models improved radiologists’ AUC from 0.929 to 0.960. At the radiologists’ sensitivity levels, the hybrid models increased their average specificity from 80.7% to 88.4% (p < 0.001), increased their PPV from 27.1% to 39.2% (p < 0.001), and decreased their average biopsy rate from 24.3% to 17.2% (p < 0.001). The reduction in biopsies using the hybrid models represented 29.4% of all recommended biopsies.*Conclusions Our AI system detected and diagnosed cancer on breast US with accuracy that exceeds that of board-certified radiologists. Our hybrid decision-making models may potentially enhance the performance of breast imagers without the added cost of a second human reader.*Clinical Relevance/Application Breast US detects additional cancers when used as a supplemental screening exam, but has high false-positive rates. AI decision support improves diagnostic accuracy by decreasing unnecessary biopsies.

On a test set of 44,755 exams, the AI system achieved an AUC of 0.976 for identifying exams with malignancies. Among the 663 reader study exams, the AI system had an AUC of 0.962, which was significantly higher than the average radiologist (0.929 ± 0.018, p < 0.001). At the average radiologist’s sensitivity (90.1%), the AI system had a higher specificity (85.6% vs 80.7%, p < 0.001) and recommended fewer biopsies (19.8% vs 24.3%, p < 0.001). On average, the hybrid models improved radiologists’ AUC from 0.929 to 0.960. At the radiologists’ sensitivity levels, the hybrid models increased their average specificity from 80.7% to 88.4% (p < 0.001), increased their PPV from 27.1% to 39.2% (p < 0.001), and decreased their average biopsy rate from 24.3% to 17.2% (p < 0.001). The reduction in biopsies using the hybrid models represented 29.4% of all recommended biopsies.

Breast US detects additional cancers when used as a supplemental screening exam, but has high false-positive rates. AI decision support improves diagnostic accuracy by decreasing unnecessary biopsies.