Arterial Calcification Disappearance in Breast Imaging: A Key Indicator for Transition to Invasive Ductal Carcinoma.

A woman in her 70s, initially suspected of having fibroadenoma due to a well-defined mass in her breast, underwent regular mammography and ultrasound screenings. Over several years, no appreciable alterations in the mass were observed, maintaining the fibroadenoma diagnosis. However, in the fourth year, an ultrasound indicated slight enlargement and peripheral irregularities in the mass, even though the mammography images at that time showed no alterations. Interestingly, mammography images over time showed the gradual disappearance of previously observed arterial calcification around the mass. Pathological examination eventually identified the mass as invasive ductal carcinoma. Although the patient had breast tissue arterial calcification typical of atherosclerosis, none was present around the tumor-associated arteries. This case highlights the importance of monitoring arterial calcification changes in mammography, suggesting that they are crucial indicators in breast cancer diagnosis, beyond observing size and shape alterations.

Investigation of imaging features in contrast-enhanced magnetic resonance imaging of benign and malignant breast lesions.

PURPOSE: This study aimed to enhance the diagnostic accuracy of contrast-enhanced breast magnetic resonance imaging (MRI) using gadobutrol for differentiating benign breast lesions from malignant ones. Moreover, this study sought to address the limitations of current imaging techniques and criteria based on the Breast Imaging Reporting and Data System (BI-RADS). MATERIALS AND METHODS: In a multicenter retrospective study conducted in Japan, 200 women were included, comprising 100 with benign lesions and 100 with malignant lesions, all classified under BI-RADS categories 3 and 4. The MRI protocol included 3D fast gradient echo T1- weighted images with fat suppression, with gadobutrol as the contrast agent. The analysis involved evaluating patient and lesion characteristics, including age, size, location, fibroglandular tissue, background parenchymal enhancement (BPE), signal intensity, and the findings of mass and non-mass enhancement. In this study, univariate and multivariate logistic regression analyses were performed, along with decision tree analysis, to identify significant predictors for the classification of lesions. RESULTS: Differences in lesion characteristics were identified, which may influence malignancy risk. The multivariate logistic regression model revealed age, lesion location, shape, and signal intensity as significant predictors of malignancy. Decision tree analysis identified additional diagnostic factors, including lesion margin and BPE level. The decision tree models demonstrated high diagnostic accuracy, with the logistic regression model showing an area under the curve of 0.925 for masses and 0.829 for non-mass enhancements. CONCLUSION: This study underscores the importance of integrating patient age, lesion location, and BPE level into the BI-RADS criteria to improve the differentiation between benign and malignant breast lesions. This approach could minimize unnecessary biopsies and enhance clinical decision-making in breast cancer diagnostics, highlighting the effectiveness of gadobutrol in breast MRI evaluations.

Deep Learning-Based Image Quality Improvement in Digital Positron Emission Tomography for Breast Cancer.

We investigated whether 18F-fluorodeoxyglucose positron emission tomography (PET)/computed tomography images restored via deep learning (DL) improved image quality and affected axillary lymph node (ALN) metastasis diagnosis in patients with breast cancer. Using a five-point scale, two readers compared the image quality of DL-PET and conventional PET (cPET) in 53 consecutive patients from September 2020 to October 2021. Visually analyzed ipsilateral ALNs were rated on a three-point scale. The standard uptake values SUVmax and SUVpeak were calculated for breast cancer regions of interest. For "depiction of primary lesion", reader 2 scored DL-PET significantly higher than cPET. For "noise", "clarity of mammary gland", and "overall image quality", both readers scored DL-PET significantly higher than cPET. The SUVmax and SUVpeak for primary lesions and normal breasts were significantly higher in DL-PET than in cPET (p < 0.001). Considering the ALN metastasis scores 1 and 2 as negative and 3 as positive, the McNemar test revealed no significant difference between cPET and DL-PET scores for either reader (p = 0.250, 0.625). DL-PET improved visual image quality for breast cancer compared with cPET. SUVmax and SUVpeak were significantly higher in DL-PET than in cPET. DL-PET and cPET exhibited comparable diagnostic abilities for ALN metastasis.

Clinical Utility of Breast Ultrasound Images Synthesized by a Generative Adversarial Network.

BACKGROUND AND OBJECTIVES: This study compares the clinical properties of original breast ultrasound images and those synthesized by a generative adversarial network (GAN) to assess the clinical usefulness of GAN-synthesized images. MATERIALS AND METHODS: We retrospectively collected approximately 200 breast ultrasound images for each of five representative histological tissue types (cyst, fibroadenoma, scirrhous, solid, and tubule-forming invasive ductal carcinomas) as training images. A deep convolutional GAN (DCGAN) image-generation model synthesized images of the five histological types. Two diagnostic radiologists (reader 1 with 13 years of experience and reader 2 with 7 years of experience) were given a reading test consisting of 50 synthesized and 50 original images (≥1-month interval between sets) to assign the perceived histological tissue type. The percentages of correct diagnoses were calculated, and the reader agreement was assessed using the kappa coefficient. RESULTS: The synthetic and original images were indistinguishable. The correct diagnostic rates from the synthetic images for readers 1 and 2 were 86.0% and 78.0% and from the original images were 88.0% and 78.0%, respectively. The kappa values were 0.625 and 0.650 for the synthetic and original images, respectively. The diagnoses made from the DCGAN synthetic images and original images were similar. CONCLUSION: The DCGAN-synthesized images closely resemble the original ultrasound images in clinical characteristics, suggesting their potential utility in clinical education and training, particularly for enhancing diagnostic skills in breast ultrasound imaging.

Comparison of 18F-fluorothymidine Positron Emission Tomography/Computed Tomography and 18F-fluorodeoxyglucose Positron Emission Tomography/Computed Tomography in Patients with Breast Cancer.

The uptake of 18F-fluorothymidine (18F-FLT) depends on cells' proliferative rates. We compared the characteristics of 18F-FLT positron emission tomography/computed tomography (PET/CT) with those of 18F-fluorodeoxyglucose (18F-FDG) PET/CT for breast cancer. We prospectively diagnosed patients with breast cancer who underwent 18F-FLT PET/CT and 18F-FDG PET/CT. Subsequently, significant differences and correlation coefficients of the maximum standardized uptake value (SUVmax) in primary breast cancer and axillary lymph nodes were statistically evaluated. We enrolled eight patients with breast cancer. In six treatment-naive patients, the SUVmax for primary lesions showed a significant difference (mean, 2.1 vs. 4.1, p = 0.031) and a strong correlation (r = 0.969) between 18F-FLT and 18F-FDG. Further, although the SUVmax for the axillary lymph nodes did not show a significant difference between 18F-FLT and 18F-FDG (P = 0.246), there was a strong correlation between the two (r = 0.999). In a patient-by-patient study, there were cases in which only 18F-FDG uptake was observed in lymph nodes and normal breasts. Bone metastases demonstrated lower accumulation than bone marrow on the 18F-FLT PET/CT. In conclusion, a strong correlation was observed between the 18F-FLT PET/CT and 18F-FDG PET/CT uptake. Differences in the biochemical characteristics of 18F-FLT and 18F-FDG were reflected in the accumulation differences for breast cancer, metastatic lesions, and normal organs.

Virtual Navigator Real-Time Ultrasound Fusion Imaging with Positron Emission Tomography/Computed Tomography for Preoperative Breast Cancer.

We used virtual navigator real-time ultrasound (US) fusion imaging with 18F-fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) to identify a lesion that could not be detected on the US alone in a preoperative breast cancer patient. Of the patient's two lesions of breast cancer, the calcified lesion could not be identified by US alone. By fusing US with 18F-FDG PET/CT, which had been performed in advance, the location of the lesion could be estimated and marked, which benefited planning an appropriate surgery. The fusion of US and 18F-FDG PET/CT was a simple and noninvasive method for identifying the lesions detected by 18F-FDG PET/CT.

Relationship between Prognostic Stage in Breast Cancer and Fluorine-18 Fluorodeoxyglucose Positron Emission Tomography/Computed Tomography.

This retrospective study examined the relationship between the standardized uptake value max (SUVmax) of fluorine-18 fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) and the prognostic stage of breast cancer. We examined 358 breast cancers in 334 patients who underwent 18F-FDG PET/CT for initial staging between January 2016 and December 2019. We extracted data including SUVmax of 18F-FDG PET and pathological biomarkers, including estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), and nuclear grade. Anatomical and prognostic stages were determined per the American Joint Committee on Cancer (eighth edition). We examined whether there were statistical differences in SUVmax between each prognostic stage. The mean SUVmax values for clinical prognostic stages were as follow: stage 0, 2.2 ± 1.4; stage IA, 2.6 ± 2.1; stage IB, 4.2 ± 3.5; stage IIA, 5.2 ± 2.8; stage IIB, 7.7 ± 6.7; and stage III + IV, 7.0 ± 4.5. The SUVmax values for pathological prognostic stages were as follows: stage 0, 2.2 ± 1.4; stage IA, 2.8 ± 2.2; stage IB, 5.4 ± 3.6; stage IIA, 6.3 ± 3.1; stage IIB, 9.2 ± 7.5, and stage III + IV, 6.2 ± 5.2. There were significant differences in mean SUVmax between clinical prognostic stage 0 and ≥II (p < 0.001) and I and ≥II (p < 0.001). There were also significant differences in mean SUVmax between pathological prognostic stage 0 and ≥II (p < 0.001) and I and ≥II (p < 0.001). In conclusion, mean SUVmax increased with all stages up to prognostic stage IIB, and there were significant differences between several stages. The SUVmax of 18F-FDG PET/CT may contribute to prognostic stage stratification, particularly in early cases of breast cancers.

Clinical usefulness of the fast protocol of breast diffusion-weighted imaging using 3T magnetic resonance imaging with a 16-channel breast coil.

We aimed to evaluate the usefulness of a fast protocol of diffusion-weighted imaging (DWI) with one excitation using 3T magnetic resonance imaging (MRI) and a 16-channel breast coil. We analyzed 30 lesions from 27 women between February 2020 and June 2020. The visibility score (from 1 = extremely poor to 5 = excellent) and apparent diffusion coefficient (ADC) value between one and four excitations were evaluated by two readers. The image acquisition time was 40 s for one excitation and 1 min 52 s for four excitations. The visibility scores were 4.630 ± 0.718 and 4.267 ± 1.015 for one excitation and 4.730 ± 0.691 and 4.200 ± 1.000 for four excitations by the two readers. There was no significant difference in the visibility (P = 0.184 and P = 0.423), mean ADC value (P = 0.918 and P = 0.417), and minimum ADC value (P = 0.936 and P = 0.443) between one and four excitations by the two readers. Despite the short acquisition time, the visibility score and ADC values of one-excitation DWI were comparable to that with four excitations. Our fast DWI protocol could provide reproducible visibility and ADC value, potentially helping radiologists to efficiently diagnose patients.

Investigating the Image Quality and Utility of Synthetic MRI in the Breast.

PURPOSE: Synthetic MRI reconstructs multiple sequences in a single acquisition. In the present study, we aimed to compare the image quality and utility of synthetic MRI with that of conventional MRI in the breast. METHODS: We retrospectively collected the imaging data of 37 women (mean age: 55.1 years; range: 20-78 years) who had undergone both synthetic and conventional MRI of T2-weighted, T1-weighted, and fat-suppressed (FS)-T2-weighted images. Two independent breast radiologists evaluated the overall image quality, anatomical sharpness, contrast between tissues, image homogeneity, and presence of artifacts of synthetic and conventional MRI on a 5-point scale (5 = very good to 1 = very poor). The interobserver agreement between the radiologists was evaluated using weighted kappa. RESULTS: For synthetic MRI, the acquisition time was 3 min 28 s. On the 5-point scale evaluation of overall image quality, although the scores of synthetic FS-T2-weighted images (4.01 ± 0.56) were lower than that of conventional images (4.95 ± 0.23; P < 0.001), the scores of synthetic T1- and T2-weighted images (4.95 ± 0.23 and 4.97 ± 0.16) were comparable with those of conventional images (4.92 ± 0.27 and 4.97 ± 0.16; P = 0.484 and 1.000, respectively). The kappa coefficient of conventional MRI was fair (0.53; P < 0.001), and that of conventional MRI was fair (0.46; P < 0.001). CONCLUSION: The image quality of synthetic T1- and T2-weighted images was similar to that of conventional images and diagnostically acceptable, whereas the quality of synthetic T2-weighted FS images was inferior to conventional images. Although synthetic MRI images of the breast have the potential to provide efficient image diagnosis, further validation and improvement are required for clinical application.

Deep-learning approach with convolutional neural network for classification of maximum intensity projections of dynamic contrast-enhanced breast magnetic resonance imaging.

PURPOSE: We aimed to evaluate deep learning approach with convolutional neural networks (CNNs) to discriminate between benign and malignant lesions on maximum intensity projections of dynamic contrast-enhanced breast magnetic resonance imaging (MRI). METHODS: We retrospectively gathered maximum intensity projections of dynamic contrast-enhanced breast MRI of 106 benign (including 22 normal) and 180 malignant cases for training and validation data. CNN models were constructed to calculate the probability of malignancy using CNN architectures (DenseNet121, DenseNet169, InceptionResNetV2, InceptionV3, NasNetMobile, and Xception) with 500 epochs and analyzed that of 25 benign (including 12 normal) and 47 malignant cases for test data. Two human readers also interpreted these test data and scored the probability of malignancy for each case using Breast Imaging Reporting and Data System. Sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC) were calculated. RESULTS: The CNN models showed a mean AUC of 0.830 (range, 0.750-0.895). The best model was InceptionResNetV2. This model, Reader 1, and Reader 2 had sensitivities of 74.5%, 72.3%, and 78.7%; specificities of 96.0%, 88.0%, and 80.0%; and AUCs of 0.895, 0.823, and 0.849, respectively. No significant difference arose between the CNN models and human readers (p > 0.125). CONCLUSION: Our CNN models showed comparable diagnostic performance in differentiating between benign and malignant lesions to human readers on maximum intensity projection of dynamic contrast-enhanced breast MRI.

Virtual Interpolation Images of Tumor Development and Growth on Breast Ultrasound Image Synthesis With Deep Convolutional Generative Adversarial Networks.

OBJECTIVES: We sought to generate realistic synthetic breast ultrasound images and express virtual interpolation images of tumors using a deep convolutional generative adversarial network (DCGAN). METHODS: After retrospective selection of breast ultrasound images of 528 benign masses, 529 malignant masses, and 583 normal breasts, 20 synthesized images of each were generated by the DCGAN. Fifteen virtual interpolation images of tumors were generated by changing the value of the input vector. A total of 60 synthesized images and 20 virtual interpolation images were evaluated by 2 readers, who scored them on a 5-point scale (1, very good; to 5, very poor) and then answered whether the synthesized image was benign, malignant, or normal. RESULTS: The mean score of overall quality for synthesized images was 3.05, and that of the reality of virtual interpolation images was 2.53. The readers classified the generated images with a correct answer rate of 92.5%. CONCLUSIONS: A DCGAN can generate high-quality synthetic breast ultrasound images of each pathologic tissue and has the potential to create realistic virtual interpolation images of tumor development.

The Utility of Deep Learning in Breast Ultrasonic Imaging: A Review.

Breast cancer is the most frequently diagnosed cancer in women; it poses a serious threat to women's health. Thus, early detection and proper treatment can improve patient prognosis. Breast ultrasound is one of the most commonly used modalities for diagnosing and detecting breast cancer in clinical practice. Deep learning technology has made significant progress in data extraction and analysis for medical images in recent years. Therefore, the use of deep learning for breast ultrasonic imaging in clinical practice is extremely important, as it saves time, reduces radiologist fatigue, and compensates for a lack of experience and skills in some cases. This review article discusses the basic technical knowledge and algorithms of deep learning for breast ultrasound and the application of deep learning technology in image classification, object detection, segmentation, and image synthesis. Finally, we discuss the current issues and future perspectives of deep learning technology in breast ultrasound.

Efficient Anomaly Detection with Generative Adversarial Network for Breast Ultrasound Imaging.

We aimed to use generative adversarial network (GAN)-based anomaly detection to diagnose images of normal tissue, benign masses, or malignant masses on breast ultrasound. We retrospectively collected 531 normal breast ultrasound images from 69 patients. Data augmentation was performed and 6372 (531 × 12) images were available for training. Efficient GAN-based anomaly detection was used to construct a computational model to detect anomalous lesions in images and calculate abnormalities as an anomaly score. Images of 51 normal tissues, 48 benign masses, and 72 malignant masses were analyzed for the test data. The sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) of this anomaly detection model were calculated. Malignant masses had significantly higher anomaly scores than benign masses (p < 0.001), and benign masses had significantly higher scores than normal tissues (p < 0.001). Our anomaly detection model had high sensitivities, specificities, and AUC values for distinguishing normal tissues from benign and malignant masses, with even greater values for distinguishing normal tissues from malignant masses. GAN-based anomaly detection shows high performance for the detection and diagnosis of anomalous lesions in breast ultrasound images.

Feasibility of new fat suppression for breast MRI using pix2pix.

PURPOSE: To generate and evaluate fat-saturated T1-weighted (FST1W) image synthesis of breast magnetic resonance imaging (MRI) using pix2pix. MATERIALS AND METHODS: We collected pairs of noncontrast-enhanced T1-weighted an FST1W images of breast MRI for training data (2112 pairs from 15 patients), validation data (428 pairs from three patients), and test data (90 pairs from 30 patients). From the original images, 90 synthetic images were generated with 50, 100, and 200 epochs using pix2pix. Two breast radiologists evaluated the synthetic images (from 1 = excellent to 5 = very poor) for quality of fat suppression, anatomic structures, artifacts, etc. The average score was analyzed for each epoch and breast density. RESULTS: The synthetic images were scored from 2.95 to 3.60; the best was reduction in artifacts when using 100 epochs. The average overall quality scores for fat suppression were 3.63 at 50 epochs, 3.24 at 100 epochs, and 3.12 at 200 epochs. In the analysis for breast density, each score was significantly better for nondense breasts than for dense breasts; the average score was 2.88-3.18 for nondense breasts and 3.03-3.42 for dense breasts (P = 0.000-0.042). CONCLUSION: Pix2pix had the potential to generate FST1W synthesis for breast MRI.

Classification of Breast Masses on Ultrasound Shear Wave Elastography using Convolutional Neural Networks.

We aimed to use deep learning with convolutional neural networks (CNNs) to discriminate images of benign and malignant breast masses on ultrasound shear wave elastography (SWE). We retrospectively gathered 158 images of benign masses and 146 images of malignant masses as training data for SWE. A deep learning model was constructed using several CNN architectures (Xception, InceptionV3, InceptionResNetV2, DenseNet121, DenseNet169, and NASNetMobile) with 50, 100, and 200 epochs. We analyzed SWE images of 38 benign masses and 35 malignant masses as test data. Two radiologists interpreted these test data through a consensus reading using a 5-point visual color assessment (SWEc) and the mean elasticity value (in kPa) (SWEe). Sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were calculated. The best CNN model (which was DenseNet169 with 100 epochs), SWEc, and SWEe had a sensitivity of 0.857, 0.829, and 0.914 and a specificity of 0.789, 0.737, and 0.763 respectively. The CNNs exhibited a mean AUC of 0.870 (range, 0.844-0.898), and SWEc and SWEe had an AUC of 0.821 and 0.855. The CNNs had an equal or better diagnostic performance compared with radiologist readings. DenseNet169 with 100 epochs, Xception with 50 epochs, and Xception with 100 epochs had a better diagnostic performance compared with SWEc (P = 0.018-0.037). Deep learning with CNNs exhibited equal or higher AUC compared with radiologists when discriminating benign from malignant breast masses on ultrasound SWE.

Breast Ultrasound Image Synthesis using Deep Convolutional Generative Adversarial Networks.

Deep convolutional generative adversarial networks (DCGANs) are newly developed tools for generating synthesized images. To determine the clinical utility of synthesized images, we generated breast ultrasound images and assessed their quality and clinical value. After retrospectively collecting 528 images of 144 benign masses and 529 images of 216 malignant masses in the breasts, synthesized images were generated using a DCGAN with 50, 100, 200, 500, and 1000 epochs. The synthesized (n = 20) and original (n = 40) images were evaluated by two radiologists, who scored them for overall quality, definition of anatomic structures, and visualization of the masses on a five-point scale. They also scored the possibility of images being original. Although there was no significant difference between the images synthesized with 1000 and 500 epochs, the latter were evaluated as being of higher quality than all other images. Moreover, 2.5%, 0%, 12.5%, 37.5%, and 22.5% of the images synthesized with 50, 100, 200, 500, and 1000 epochs, respectively, and 14% of the original images were indistinguishable from one another. Interobserver agreement was very good (|r| = 0.708-0.825, p < 0.001). Therefore, DCGAN can generate high-quality and realistic synthesized breast ultrasound images that are indistinguishable from the original images.

Diagnostic performance of time-of-flight PET/CT for evaluating nodal metastasis of the axilla in breast cancer.

OBJECTIVE: The aim of this study was to evaluate the performance of preoperative axillary lymph node assessment in breast cancer using time-of-flight 18F-fluorodeoxyglucose PET/computed tomography (TOF [F-18]FDG-PET/CT). METHODS: Eighty-two women with breast cancer (mean age, 59.3 years; range, 30-84 years) underwent TOF [F-18]FDG-PET/CT scanning before surgery between January 2016 and June 2018 at our hospital. Visual analysis of FDG uptake and the maximum standardized uptake value (SUVmax) of axillary lymph nodes were compared with the pathological diagnoses. RESULTS: There were 77 patients with invasive breast carcinoma (mean invasive long diameter, 18.5 mm; range, 2-90 mm) and five patients with noninvasive carcinoma. Axillary lymph node metastases were histologically confirmed in 13 of 82 patients (15.9%). SUVmax showed an area under a receiver operating characteristic curve of 0.916, and the cut-off value of 1.1 was appropriate. By visual assessment, there were 11 true positives, 15 false positives, 54 true negatives and two false negatives; the sensitivity, specificity, positive predictive value, negative predictive value and accuracy were 85%, 78%, 42%, 96% and 79%, respectively. SUVmax showed values of 69%, 99%, 90%, 94% and 94%, respectively. CONCLUSIONS: The sensitivity of TOF [F-18]FDG-PET/CT was as high as 85% by visual analysis. SUVmax using TOF [F-18]FDG-PET/CT showed high diagnostic performance for N-staging in breast cancer patients, especially high negative predictive value. The specificity, positive predictive value and accuracy of SUVmax were higher than those of visual analysis.

Simultaneous comparison between strain and shear wave elastography of breast masses for the differentiation of benign and malignant lesions by qualitative and quantitative assessments.

PURPOSE: To compare the addition of diagnostic strain elastography (SE) and shear wave elastography (SWE) values to the conventional B-mode ultrasonography in differentiating between benign and malignant breast masses by qualitative and quantitative assessments. MATERIALS AND METHODS: B-mode ultrasound, SE, and SWE were simultaneously performed using one ultrasound system in 148 breast masses; 88 of them were malignant. The breast imaging reporting and data system category in the B-mode, Tsukuba score (SETsu), Fat-Lesion-Ratio (SEFLR) in SE, and five-point color assessment (SWEcol) and elasticity values (SWEela) in SWE were assessed. The results were compared using the area under the receiver-operating characteristic curve (AUC). RESULT: The AUC for B-mode and each elastography were similar (B-mode, 0.889; SETsu, 0.885; SEFLR, 0.875; SWEcol, 0.881; SWEela, 0.885; P > 0.05). The combined sets between B-mode and either of the elastography technique showed good diagnostic performance (B-mode + SETsu, 0.903; B-mode + SEFLR, 0.909; B-mode + SWEcol, 0.919; B-mode + SWEela, 0.914). B-mode + SWEcol and B-mode + SWEela showed a higher AUC than B-mode alone (P = 0.026 and 0.029), and B-mode + SETsu and B-mode + SEFLR showed comparable AUC to B-mode alone (P = 0.196 and 0.085). There was no significant difference between qualitative and quantitative assessments for the combined sets of B-mode and elastography (P > 0.05). CONCLUSION: The addition of both SE and SWE to B-mode ultrasound improved the diagnostic performance with increased AUC, and especially SWE was more useful than SE, and no significant difference was found between qualitative and quantitative assessments.

Distinction between benign and malignant breast masses at breast ultrasound using deep learning method with convolutional neural network.

PURPOSE: We aimed to use deep learning with convolutional neural network (CNN) to discriminate between benign and malignant breast mass images from ultrasound. MATERIALS AND METHODS: We retrospectively gathered 480 images of 96 benign masses and 467 images of 144 malignant masses for training data. Deep learning model was constructed using CNN architecture GoogLeNet and analyzed test data: 48 benign masses, 72 malignant masses. Three radiologists interpreted these test data. Sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC) were calculated. RESULTS: The CNN model and radiologists had a sensitivity of 0.958 and 0.583-0.917, specificity of 0.925 and 0.604-0.771, and accuracy of 0.925 and 0.658-0.792, respectively. The CNN model had equal or better diagnostic performance compared to radiologists (AUC = 0.913 and 0.728-0.845, p = 0.01-0.14). CONCLUSION: Deep learning with CNN shows high diagnostic performance to discriminate between benign and malignant breast masses on ultrasound.