YOLO AI model based on an automated breast volume scanner for the detection of benign and malignant breast lesions

Introduction

In 2022, there were 20 million new cases of cancer, of which female breast cancer accounted for 11.6% of the total (1). Breast cancer is the most common cancer in women in terms of both incidence and mortality (1). However, mortality and morbidity vary across countries due to factors such as limited access to early screening, timely detection, and cost-effective treatment (1,2). Research has shown that women who participate in breast cancer screening programs have a 60% and 47% lower risk of dying from breast cancer within 10 and 20 years of diagnosis, respectively (3). Thus, the early detection of breast cancer is critical for reducing mortality rates (4).

Breast ultrasound is a frequently used diagnostic modality in clinical medicine (5). This technique has increased sensitivity in the identification of breast cancer among high-risk individuals with dense breast tissue (6-8). However, operation and diagnosis using conventional ultrasonography are highly operator-dependent and less reproducible. Recently, the automated breast volume scanner (ABVS) has led to new breakthroughs in the diagnosis of breast diseases due to its improved operator independence and reproducibility (9,10). The ABVS has been reported to show a high degree of agreement with conventional ultrasound in the measurement of lesion size [intraclass correlation coefficient (ICC): 0.95] and a higher agreement with histology than conventional ultrasound (ICC: 0.85 and 0.75, respectively) (11). Further, in the detection of smaller invasive carcinomas, the ABVS has shown an advantage over mammography (12). However, due to the large number of images in ABVS datasets, lesion detection and diagnosis are time-consuming.

Recently, the rapid rise of artificial intelligence (AI) technology, particularly deep learning, has resulted in significant advances in the detection and diagnosis of breast lesions (13-16). Chiang et al. (17) used a three-dimensional (3D) convolutional neural network for lesion detection in 171 test sets, and reported sensitivities of 95%, 90%, 85%, and 80% at 14.03, 6.92, 4.91, and 3.62 false positive (FP) lesions per patient, respectively. Using an improved deep-learning model, Inception V3, Zhang et al. (18) achieved an 88.6% accuracy in the diagnosis of breast cancer in 149 breast nodules. However, the small sample sizes of these studies did not allow for a simultaneous focus on lesion detection and the diagnosis of benign and malignant properties, which limits the validity of such models in clinical applications.

Thus, this study used a substantial breast dataset to construct an ABVS detection model employing the neural network framework of You Only Look Once (YOLO; version 8) to alleviate the burden placed on radiologists and to enhance breast cancer detection rates. Further, it analyzed the performance of the model in detecting breast lesions of varying sizes, and assessed changes in the diagnostic performance of radiologists before and after the integration of the model. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-24-2079/rc).

Methods

Study population

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Medical Ethics Committee of The First Affiliated Hospital of Anhui Medical University (Approval No. PJ 2024-02-12) and The Second People’s Hospital of Wuhu (Approval No. 2024-KY-103), and the requirement of individual consent for this retrospective analysis was waived.

The data of 1,284 patients who underwent ABVS examination between January 2017 and August 2021 at The Second People’s Hospital of Wuhu were retrospectively, consecutively, and unselectively collected.

Patients were included in the study if they met the following inclusion criteria: (I) had puncture biopsy or surgical pathological results of the target lesion obtained within two weeks of ultrasonography; (II) had preserved, intact ABVS full-volume images; and (III) had undergone target lesion sonograms.

Patients were excluded from the study if they met any of the following exclusion criteria: (I) had images containing multiple lesions with some lesions lacking pathological results; and/or (II) had borderline tumors, such as borderline phyllodes tumors.

The patient enrollment flowchart for the study is shown in Figure 1.

Figure 1 Flowchart of patient enrollment. ABVS, automated breast volume scanner.

Collection of clinical data

The baseline data of the patients, including their age and pathology results, were obtained from the hospital’s information system. The maximum lesion diameter was determined using ultrasound reports. Each lesion was diagnosed based on the histological examination of a biopsy or surgical specimen. The lesions were classified as benign or malignant based on the pathological findings.

ABVS data acquisition

The ABVS images were obtained using the Acuson S2000 ABVS (Siemens Medical Solutions, Erlangen, Germany). The ABVS was equipped with a 14L5BV (5–14 MHz) automatically driven linear array probe. Anteroposterior and lateral views were obtained for each breast. Medial, inferior, superior, and axillary images were acquired as necessary.

Data preprocessing and labeling

To aid in lesion detection and analysis, the ABVS datasets were converted into continuous two-dimensional (2D) images, each comprising 318 individual images. A radiologist with 11 years of experience screened the 2D images of each target lesion from each dataset, spanning from lesion emergence to maximum lesion size and lesion disappearance (Figure S1). Subsequently, the radiologist used “LabelImg” to manually label the lesion in each image with a rectangular box that included the entire lesion. These rectangular boxes marked the location information of the mass bounding box, and were saved as Extensible Markup Language (XML) files. The attribute tags (benign or malignant) were then uniformly added to each XML file by the radiologist using Python software (version 3.7; Python Software Foundation, Wilmington, DE, USA) based on the pathology of the lesion. After two weeks, 120 lesions (1,923 2D images) were randomly re-labeled by the same radiologist and another radiologist with eight years of ultrasound experience. Intra- and inter-observer bounding box consistency was assessed using the intersection-over-union (IoU) ratio. The radiologists were blinded to the pathology results of the patients.

Construction of the YOLO AI model

All the data were randomly partitioned according to each lesion, and divided into a training set and a test set at a ratio of 7 to 3. The training set was used for model training, and the test set was used for model testing. The models were tri-fold cross-validated, and the optimal model was selected and analyzed. In this study, a YOLO AI model was established for target detection, and the classification algorithm was based on the YOLO (version 8) model (19). YOLO AI comprises the following three key components: a channel separable convolution (CSP)-enhanced Darknet53 backbone, a multi-scale feature fusion neck, and an anchor-free detection head. The framework of the network model is shown in Figure 2, and the detailed structure of the YOLO AI model is shown in Appendix 1. The images were subjected to preprocessing operations before training, and then input into the YOLO AI model; details of the training methodology and process are provided in Appendix 2 and Figure S2.

Figure 2 Overall framework of the YOLO AI model. AI, artificial intelligence; CSP, channel separable convolution; YOLO, You Only Look Once.

Evaluation criteria for the YOLO AI models

In this study, the performance of the YOLO AI model was evaluated in terms of its precision, recall, mean average precision (mAP)50, and F1-score using the test set. Precision refers to how accurately a model predicts a positive example. Recall refers to the ability of a model to recognize a positive example. For each category of detection results, the precision and recall were obtained and plotted as precision-recall (P-R) curves. The average precision (AP) refers to the area enclosed by the P-R curve and x-axis, which was a measure of how well the model performed in the detection of this category. mAP50 is the mAP calculated with an IoU threshold of 0.5, which was a measure of the model’s target detection performance. The following formulas were used to calculate precision, recall, the mAP, and the F1-score, respectively:

P(Precision)=TPTP+FP[1]

Recall=TPTP+FN[2]

mAP=1N∑i=1NAPi[3]

F1-score=2×P×RecallP+Recall[4]

where TP represents true positive, FP represents false positive, FN represents false negative, N represents the total number of categories, and AP_i represents the AP for the i-th category.

The IoU was calculated using the following formula:

IoU=Boundingbox∩GoundtruthBoundingbox∪Goundtruth[5]

Additionally, the test set data were divided into four groups based on the diameter of the lesions (diameter ≤10 mm, 10 mm < diameter ≤ 20 mm, 20 mm < diameter ≤ 30 mm, and diameter >30 mm), and the performance of the YOLO AI model was then evaluated in each of these four groups. In this part of the study, 26 images containing multiple lesions were excluded because these lesions were categorized differently in terms of diameter.

Comparison of the diagnostic efficacy of YOLO AI model-assisted radiologists with different years of experience

Readers 1, 2, 3, and 4 (radiologists with 1, 5, 10, and 18 years of experience, respectively) reviewed all the images, and categorized each lesion as either benign or malignant based on their experience. Two weeks later, the same four radiologists re-evaluated all the lesions, using the YOLO AI results to re-diagnose the lesions. The effectiveness of the radiologists in diagnosing the breast lesions before and after using the model was evaluated based on the area under the curve (AUC) of the receiver operating characteristic (ROC) curve. The radiologist was not involved in the annotation of the images and was blinded to the pathological results.

Statistical analysis

The statistical analysis was performed using R software 4.1.2 (R Foundation for Statistical Computing, Vienna, Austria) and SPSS 25.0 software (IBM, Armonk, NY, USA). The continuous variables are expressed as the mean ± standard deviation, or the median [interquartile range (IQR)]. The categorical variables are presented as the count or percentage. Differences in the continuous variables between the groups were assessed using the Mann-Whitney U test or Student’s t-test. Differences in the categorical variables between the groups were assessed using the chi-squared test (χ²) or Fisher’s test. DeLong’s test was used to assess the difference in the AUC values before and after the radiologist’s combined use of the YOLO AI model. A two-sided P<0.05 was considered statistically significant.

Results

Patient and lesion characteristics

A total of 1,284 patients with 1,769 lesions were included in this study. Of these lesions, 1,296 (73.3%) were benign, and 473 (26.7%) were malignant. The lesions were randomly divided into training and test sets at a 7:3 ratio. The training set comprised 1,242 (70.2%) lesions, and the test set comprised 527 (29.8%) lesions. There were 3–95 2D images per lesion for a total of 26,099 images. The number of lesions in each image ranged from 1 to 2, totaling 26,374 2D lesions. The distribution of the baseline information of the patients in this study is presented in Table 1. There were no statistically significant differences between the training and test sets in terms of the distribution of lesion location, swept view, number of lesions, number of 2D lesions, lesion diameter, and patient age (P>0.05). The distribution of the pathological data for all lesions is shown in Table S1. The data were divided according to the optimal model in the tri-fold cross-validation results during the model training process. The results of the tri-fold cross-validation are provided in Table S2 and Figure S3.

Good intra- and inter-observer reproducibility was observed, with IoU IQR values of 0.869 (0.106) and 0.857 (0.103), respectively. The intra-observer IoU values were 0.869 (0.103) and 0.870 (0.116) for the benign and malignant lesions, respectively, with no statistically significant difference (P=0.644). The inter-observer IoU values were 0.859 (0.103) and 0.851 (0.104) for the benign and malignant lesions, respectively, with no statistically significant difference (P=0.173).

Efficacy of the YOLO AI model in single-target detection

Figure 3A shows the confusion matrix of the YOLO AI model for the single-target detection of the images (i.e., distinguishing between the background and lesions) in the test set. The recall rate of the model for single-target detection was 98.3% (7,986/8,125). The recall rates of the YOLO AI model for benign and malignant 2D lesions were 99.1% (2,591/2,615) and 97.9%, respectively, with statistically significant differences (P<0.001). An example of single-target detection using the YOLO AI model is provided in Figure S4.

Figure 3 Confusion matrix of the YOLO AI model. (A) Single-target detection; (B) benign and malignant lesion detection; (C-F) benign and malignant lesions with a diameter ≤10 mm (C), 10 mm < diameter ≤ 20 mm (D), 20 mm < diameter ≤ 30 mm (E), or diameter >30 mm (F). AI, artificial intelligence; YOLO, You Only Look Once.

Efficacy of the YOLO AI model in classification objective detection

The efficacy of the YOLO AI model for classification objective detection (i.e., distinguishing between benign and malignant lesions) is shown in Table 2. The model had an overall precision, recall, mAP50, and F1-score of 0.887, 0.866, 0.919, and 0.876, respectively. The precision, recall, mAP50, and F1-score of the model for detecting benign and malignant lesions were 0.884, 0.927, and 0.939, and 0.905, 0.890, 0.806, 0.899, and 0.846, respectively. Figure 3B shows the confusion matrix of the YOLO AI model for benign and malignant lesion detection.

Efficacy of the YOLO AI model in detecting lesions of different diameters

The efficacy of the YOLO AI model in detecting benign and malignant lesions with a diameter ≤10 mm, 10 mm < diameter ≤ 20 mm, 20 mm < diameter ≤ 30 mm, and diameter >30 mm is shown in Table 2. The precision of the YOLO AI model (0.876–0.910) in detecting benign and malignant lesions with different diameters was high, and its precision in detecting malignant lesions was better than that of its precision in detecting benign lesions. The recall of the model increased with diameter, especially for malignant lesions. Its recall for benign lesions remained relatively stable across diameters (0.914–0.941). Figure 3C-3F shows the confounding matrix of the YOLO AI model for the classification of breast lesions of various diameters. Examples of the YOLO AI model for the detection of benign and malignant lesions of different diameters are provided in Figure 4.

Figure 4 Examples of the diagnosis of the YOLO AI model for breast lesions of different diameters. (A,B) Benign and malignant lesions with a diameter ≤10 mm. (C,D) Benign and malignant lesions with a 10 mm < diameter ≤ 20 mm. (E,F) Benign and malignant lesions with a 20 mm < diameter ≤ 30 mm. (G,H) Benign and malignant lesions with a diameter >30 mm. AI, artificial intelligence; YOLO, You Only Look Once.

Comparison of the diagnostic efficacy of YOLO AI-assisted radiologists with different years of experience

The performance results of readers 1, 2, 3, and 4 in the diagnosis of breast lesions before and after receiving YOLO AI model assistance are shown in Table 3 and Figure 5. The DeLong test showed that the AUC value after combining the models for all readers was superior to that of the AUC values for the readers’ diagnoses alone (P<0.001). In addition, there were no significant differences in the diagnostic performance between readers 2 and 3 (P=0.641), readers 2 and 4 (P=0.780), and readers 3 and 4 (P=0.917) after receiving AI assistance; however, these three readers all outperformed reader 1 (P<0.001). The difference in the AUC values between readers 1 and 4 after receiving AI assistance was not statistically significant (P=0.890). The reading times of the four readers are shown in Table 3, and the average reduction in reading time after receiving YOLO AI assistance was 45.5±24.24 min. The YOLO AI model detected each image at a rate of 0.8 ms.

Figure 5 ROC curves of breast lesions diagnosed by four readers before and after using the YOLO AI model. (A) ROC curves of the diagnoses of the four readers without the use of the YOLO AI model; (B) ROC curves of the diagnoses of the four readers after using the YOLO AI model. AI, artificial intelligence; AUC, area under the curve; FPR, false positive rate; ROC, receiver operating characteristic; TPR, true positive rate; YOLO, You Only Look Once.

Discussion

In this study, we developed a YOLO AI model for the rapid detection of breast lesions in ABVS sonograms. The model had 98.3% recall in single-target detection. It also effectively detected benign and malignant lesions with a precision, recall, mAP50, and F1-score of 0.887, 0.866, 0.919, and 0.876, respectively. Notably, the YOLO AI model showed stable performance for lesions of different diameters. Further, the model effectively assisted radiologists with all years of experience, and bridged the gap between radiologists with different levels of experience.

In clinical practice, the detection and diagnosis of breast lesions is of equal value; however, in previous studies, ABVSs have rarely been used to perform both tasks (17,18,20,21). In the present study, we used the data of 1,284 patients to build and validate a YOLO AI model, which had a high recall of 98.3% for lesions, which was slightly higher than that of 3D detection or multilayer feature detection methods constructed by other researchers (90.0% to 96.3%) (17,20-22). Only 1.7% of the 2D lesions in the study were not detected. Detection failure occurred most commonly in images in which the lesions: (I) had just appeared or were about to disappear; (II) were close to the fat layer or similar to the echoes of the fat layer; and (III) were located in the background disorder. The malignant lesions were detected at a higher rate in single-target detection; however, this might be because malignant lesions show more heterogeneity in sonograms, making detection easier. Hypoechogenicity is a common feature of tumors (23), and can be used to effectively distinguish between normal tissues and masses. Significant hypoechogenicity is considered a typical feature of malignant tumors (24) that can help models distinguish between benign and malignant lesions.

The YOLO AI model performed equally well in the classification objective, with precision, recall, mAP50, and an F1-score of 0.887, 0.866, 0.919, and 0.876, respectively. The precision and mAP50 of the YOLO AI model were similar for the benign and malignant lesions; however, the recall of the model was higher for benign lesions than malignant lesions. Comparisons with other models were not possible due to the limited number of previous studies that have used ABVS data for the categorization and detection of lesions, and the use of different private databases. Li et al. (25) established a BUSnet deep-learning model for the detection of benign and malignant breast lesions on conventional ultrasound, which had good recall for both benign and malignant lesions (1.000 and 0.880, respectively), but its precision (0.629) was lower for malignant lesions. Cao et al. (26) trained 10 deep-learning models, and the best-performing SSD300+ZFNet model had a high average precision rate (96.89%) but a low average recall (67.23%). Moreover, they used a preprocessing strategy of cropping the image to reduce background interference in their detection research; however, cropping may lead to limitations in practical applications. Conversely, the present study directly used complete ABVS images for detection, which more closely reflects clinical application scenarios. A complete image provides comprehensive information about the lesion and helps improve the accuracy and reliability of detection.

A detailed analysis of the lesion detection results for different diameters was performed to comprehensively assess the applicability of the model, and the overall accuracy of the model was high across all diameters (0.876–0.910). The recall rates remained consistently high across varying diameters (0.806–0.898). The model had a high precision rate (0.925) but a low recall rate (0.693) in the detection of malignant lesions with a diameter of ≤10 mm. This might be because there were fewer malignant lesions with a diameter ≤10 mm (75 2D lesions), which biased the results. Alternatively, it might be that while the model can detect lesions; it encountered difficulties in distinguishing between benign and malignant lesions as seen in the confusion matrix in Figure 3A. Notably, lesions are usually small if they are benign or they are in the early stages of cancer, and such lesions are often difficult to differentiate because they have similar imaging appearances. As the size of the lesion increases, the malignant mass expands and exhibits greater heterogeneity.

To further examine the value of the YOLO AI model in clinical settings, we observed the performance of radiologists before and after the incorporation of the YOLO AI model. Four radiologists with varying years of experience participated in the study. The AUC values for the diagnosis of breast lesions by radiologists with 1, 5, 11, and 18 years of ultrasound experience were 0.689, 0.705, 0.734, and 0.802, respectively. To exclude the effect of the first review, these four radiologists re-read the images with the assistance of the model after two weeks, and the AUC values of the four radiologists were 0.806, 0.890, 0.897, and 0.895, respectively. Similar to other studies, the radiologists’ diagnoses improved with the help of the AI model (18,27,28). In this study, the radiologists’ AUC values, sensitivity, specificity, and accuracy increased by more than 10%, and the mean reading time also decreased to 45.5±24.24 min. The junior radiologists achieved a similar level of diagnosis to the senior radiologists with AI assistance, and there was no significant difference in the diagnostic performance between the intermediate and senior radiologists with AI assistance.

This study had several limitations. First, it was a single-center retrospective study, which might have caused selection bias. Second, while the processing of the ABVS data into 2D images in this study increased the rate of detection, it decreased the value of the ABVS data in the coronal imaging. Third, this study only used data with pathological results, and thus did not include some small, clinically suspect benign lesions. This potential selection bias might have optimized the model’s performance. Finally, while the differences between the training and test sets were not statistically significant in the study, and the YOLO AI model showed stable diagnostic performance for lesions of different diameters, the stochastic nature of the data partitioning might have led to volatility in the performance evaluation of the model, and the analysis strategy of meritocracy might have led to an optimistic estimation of the model’s true generalization ability. The stability of the model will be improved in future studies by increasing the number of data randomization divisions.

Conclusions

The developed YOLO AI model exhibited encouraging performance in the identification of breast lesions, enabling the rapid and precise detection of their location and characteristics. The YOLO AI model showed stable diagnostic performance across a wide range of lesion diameters, providing effective support for radiologists of all experience levels. In the future, data will be collected from multiple centers to optimize the model and provide stronger support for the early detection and treatment of breast cancer.

Acknowledgments

We would like to thank The Second People’s Hospital of Wuhu City for providing data support and Hebin Intelligent Robots for providing technical support for this study.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-24-2079/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-24-2079/dss

Funding: This work was supported by the Anhui Provincial Natural Science Foundation (Grant number: 2308085MH278), the Health Research Program of Anhui (Grant number: AHWJ2023A10017), the Shenzhen National Research Institute of High-Performance Medical Devices Co., Ltd. (Grant number: 662202404012), the Specialized Research Initiation Foundation of The First Affiliated Hospital of Wannan Medical College (Grant number: YR202436) and the Anhui Provincial Health Commission Scientific Research Project (Grant number: AHWJ2024Aa30096).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-24-2079/coif). D.C. is currently affiliated with Hebin Intelligent Robots Co., Ltd. (a for-profit company), but this affiliation does not present any financial or non-financial conflicts of interest in relation to the submitted work. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Medical Ethics Committee of The First Affiliated Hospital of Anhui Medical University (Approval No. PJ 2024-02-12) and The Second People’s Hospital of Wuhu (Approval No. 2024-KY-103), and the requirement of individual consent for this retrospective analysis was waived.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Ginsburg O, Bray F, Coleman MP, Vanderpuye V, Eniu A, Kotha SR, et al. The global burden of women's cancers: a grand challenge in global health. Lancet 2017;389:847-60. [Crossref] [PubMed]
3. Tabár L, Dean PB, Chen TH, Yen AM, Chen SL, Fann JC, Chiu SY, Ku MM, Wu WY, Hsu CY, Chen YC, Beckmann K, Smith RA, Duffy SW. The incidence of fatal breast cancer measures the increased effectiveness of therapy in women participating in mammography screening. Cancer 2019;125:515-23. [Crossref] [PubMed]
4. Irfan R, Almazroi AA, Rauf HT, Damaševičius R, Nasr EA, Abdelgawad AE. Dilated Semantic Segmentation for Breast Ultrasonic Lesion Detection Using Parallel Feature Fusion. Diagnostics (Basel) 2021;11:1212. [Crossref] [PubMed]
5. Pourasad Y, Zarouri E, Salemizadeh Parizi M, Salih Mohammed A. Presentation of Novel Architecture for Diagnosis and Identifying Breast Cancer Location Based on Ultrasound Images Using Machine Learning. Diagnostics (Basel) 2021;11:1870. [Crossref] [PubMed]
6. Leconte I, Feger C, Galant C, Berlière M, Berg BV, D'Hoore W, Maldague B. Mammography and subsequent whole-breast sonography of nonpalpable breast cancers: the importance of radiologic breast density. AJR Am J Roentgenol 2003;180:1675-9. [Crossref] [PubMed]
7. Yang L, Wang S, Zhang L, Sheng C, Song F, Wang P, Huang Y. Performance of ultrasonography screening for breast cancer: a systematic review and meta-analysis. BMC Cancer 2020;20:499. [Crossref] [PubMed]
8. Madjar H. Role of Breast Ultrasound for the Detection and Differentiation of Breast Lesions. Breast Care (Basel) 2010;5:109-14. [Crossref] [PubMed]
9. Schmachtenberg C, Fischer T, Hamm B, Bick U. Diagnostic Performance of Automated Breast Volume Scanning (ABVS) Compared to Handheld Ultrasonography With Breast MRI as the Gold Standard. Acad Radiol 2017;24:954-61. [Crossref] [PubMed]
10. Li JM, Shao YH, Sun XM, Shi J. Ultrasonic features of automated breast volume scanner (ABVS) and handheld ultrasound (HHUS) combined with molecular biomarkers in predicting axillary lymph node metastasis of clinical T1-T2 breast cancer. Quant Imaging Med Surg 2024;14:1359-68. [Crossref] [PubMed]
11. Girometti R, Zanotel M, Londero V, Linda A, Lorenzon M, Zuiani C. Automated breast volume scanner (ABVS) in assessing breast cancer size: A comparison with conventional ultrasound and magnetic resonance imaging. Eur Radiol 2018;28:1000-8. [Crossref] [PubMed]
12. Mandelson MT, Oestreicher N, Porter PL, White D, Finder CA, Taplin SH, White E. Breast density as a predictor of mammographic detection: comparison of interval- and screen-detected cancers. J Natl Cancer Inst 2000;92:1081-7. [Crossref] [PubMed]
13. Kim J, Haffty BG. Genetic Factors in the Screening and Imaging for Breast Cancer. Korean J Radiol 2023;24:378-83. [Crossref] [PubMed]
14. Li H, Shi S, Lou L, Cao J, Zhou Z, Huang X, Ye J. Multidimensional quantitative characterization of periocular morphology: distinguishing esotropia from epicanthus by deep learning network. Quant Imaging Med Surg 2024;14:6273-84. [Crossref] [PubMed]
15. Zhong L, Shi L, Zhou L, Liu X, Gu L, Bai W. Development of a nomogram-based model combining intra- and peritumoral ultrasound radiomics with clinical features for differentiating benign from malignant in Breast Imaging Reporting and Data System category 3-5 nodules. Quant Imaging Med Surg 2023;13:6899-910. [Crossref] [PubMed]
16. Pan H, Shi C, Zhang Y, Zhong Z. Artificial intelligence-based classification of breast nodules: a quantitative morphological analysis of ultrasound images. Quant Imaging Med Surg 2024;14:3381-92. [Crossref] [PubMed]
17. Chiang TC, Huang YS, Chen RT, Huang CS, Chang RF. Tumor Detection in Automated Breast Ultrasound Using 3-D CNN and Prioritized Candidate Aggregation. IEEE Trans Med Imaging 2019;38:240-9. [Crossref] [PubMed]
18. Zhang P, Ma Z, Zhang Y, Chen X, Wang G. Improved Inception V3 method and its effect on radiologists' performance of tumor classification with automated breast ultrasound system. Gland Surg 2021;10:2232-45. [Crossref] [PubMed]
19. Jocher G, Chaurasia A, Qiu J. Ultralytics YOLOv8. Accessed on 18 November 2023. 2023. Available online: https://docs.ultralytics.com/models/yolov8/
20. Wang Y, Wang N, Xu M, Yu J, Qin C, Luo X, Yang X, Wang T, Li A, Ni D. Deeply-Supervised Networks With Threshold Loss for Cancer Detection in Automated Breast Ultrasound. IEEE Trans Med Imaging 2020;39:866-76. [Crossref] [PubMed]
21. Li Y, Wu W, Chen H, Cheng L, Wang S. 3D tumor detection in automated breast ultrasound using deep convolutional neural network. Med Phys 2020;47:5669-80. [Crossref] [PubMed]
22. Oh K, Lee SE, Kim EK. 3-D breast nodule detection on automated breast ultrasound using faster region-based convolutional neural networks and U-Net. Sci Rep 2023;13:22625. [Crossref] [PubMed]
23. Skaane P, Engedal K. Analysis of sonographic features in the differentiation of fibroadenoma and invasive ductal carcinoma. AJR Am J Roentgenol 1998;170:109-14. [Crossref] [PubMed]
24. Stavros AT, Thickman D, Rapp CL, Dennis MA, Parker SH, Sisney GA. Solid breast nodules: use of sonography to distinguish between benign and malignant lesions. Radiology 1995;196:123-34. [Crossref] [PubMed]
25. Li Y, Gu H, Wang H, Qin P, Wang J. BUSnet: A Deep Learning Model of Breast Tumor Lesion Detection for Ultrasound Images. Front Oncol 2022;12:848271. [Crossref] [PubMed]
26. Cao Z, Duan L, Yang G, Yue T, Chen Q. An experimental study on breast lesion detection and classification from ultrasound images using deep learning architectures. BMC Med Imaging 2019;19:51. [Crossref] [PubMed]
27. Wang Y, Choi EJ, Choi Y, Zhang H, Jin GY, Ko SB. Breast Cancer Classification in Automated Breast Ultrasound Using Multiview Convolutional Neural Network with Transfer Learning. Ultrasound Med Biol 2020;46:1119-32. [Crossref] [PubMed]
28. Hejduk P, Marcon M, Unkelbach J, Ciritsis A, Rossi C, Borkowski K, Boss A. Fully automatic classification of automated breast ultrasound (ABUS) imaging according to BI-RADS using a deep convolutional neural network. Eur Radiol 2022;32:4868-78. [Crossref] [PubMed]