Automated breast ultrasound features associated with diagnostic performance of a multiview convolutional neural network according to the level of experience of radiologists

Abstract

Purpose

To investigate automated breast ultrasound (ABUS) features affecting the use of a multiview convolutional neural network (CNN) for breast lesions according to the level of experience of radiologists.

Materials and Methods

A total of 656 breast lesions (152 malignant and 504 benign lesions) were included and reviewed by 6 radiologists for background echotexture, glandular tissue component (GTC), and lesion type and size without as well as with a multiview CNN. The sensitivity, specificity, and the area under the receiver operating curve (AUC) for ABUS features were compared between 2 sessions according to the level of the radiologists’ experience.

Results

Radiology residents showed significant AUC improvement with the multiview CNN for mass (0.81–0.91, P=0.003) and non-mass lesions (0.56–0.90, P=0.007), all background echotextures (homogeneous-fat: 0.84–0.94, P=0.04; homogeneous-fibroglandular: 0.85–0.93, P=0.01; heterogeneous: 0.68–0.88, P=0.002), all GTC levels (minimal: 0.86–0.93, P=0.001; mild: 0.82–0.94, P=0.003; moderate: 0.75–0.88, P=0.01; marked: 0.68–0.89, P<0.001), and lesions ≤10mm (≤5mm: 0.69–0.86, P<0.001; 6–10mm: 0.83–0.92, P<0.001). Breast specialists showed significant AUC improvement with the multiview CNN in heterogeneous echotexture (0.90–0.95, P=0.03), marked GTC (0.88–0.95, P<0.001), and lesions ≤10mm (≤5mm: 0.89–0.93, P=0.02; 6–10mm: 0.95–0.98, P=0.01).

Conclusion

With the multiview CNN, ABUS performance among radiology residents was improved regardless of lesion type, background echotexture, or GTC. For breast lesions smaller than 10mm, both radiology residents and breast specialists achieved better ABUS performance.

Zusammenfassung

Ziel

Untersuchung der Merkmale im automatisierten Brust-Ultraschall (ABUS), die sich auf die Verwendung eines Multiview-Convolutional-Neural-Networks (CNN) für Brustläsionen, abhängig von der Erfahrung der Radiologen, auswirken.

Materialien und Methoden

Insgesamt wurden 656 Brustläsionen (152 maligne und 504 benigne Läsionen) einbezogen und von 6 Radiologen hinsichtlich der Hintergrund-Echostruktur, der glandulären Gewebekomponente (GTC) sowie in Bezug auf Läsionstyp und -größe – sowohl mit als auch ohne Multiview-CNN – untersucht. Sensitivität, Spezifität und die Fläche unter der ROC-Kurve (AUC) für ABUS-Merkmale wurden zwischen 2 Sitzungen in Abhängigkeit von der Erfahrung der Radiologen verglichen.

Ergebnisse

Radiologie-Assistenzärzte zeigten mit dem Multiview-CNN eine signifikante Verbesserung der AUC für Raumforderungen (0,81–0,91; p=0,003) und Nicht-Raumforderungen (0,56–0,90; p=0,007) sowie für alle Hintergrund-Echostrukturen (homogen-fett: 0,84–0,94; p=0,04; homogen-fibroglandulär: 0,85–0,93; p=0,01; heterogen: 0,68–0,88; p=0,002), allen GTC-Stufen (minimal: 0,86–0,93; p=0,001; leicht: 0,82–0,94; p=0,003; mittel: 0,75–0,88; p=0,01; ausgeprägt: 0,68–0,89; p<0,001) und Läsionen ≤10mm (≤5mm: 0,69–0,86; p<0,001; 6–10mm: 0,83–0,92; p<0,001). Brustspezialisten zeigten mit dem Multiview-CNN eine signifikante Verbesserung der AUC bei heterogener Echostruktur (0,90–0,95; p=0,03), ausgeprägter GTC (0,88–0,95, p<0,001) und Läsionen ≤10mm (≤5mm: 0,89–0,93; p=0,02; 6–10mm: 0,95–0,98; p = 0,01).

Schlussfolgerung

Mit dem Multiview-CNN verbesserte sich die ABUS-Leistung von Radiologie-Assistenzärzten – unabhängig von der Art der Läsion, der Hintergrund-Echotextur oder der GTC. Bei Brustläsionen kleiner als 10mm erzielten sowohl Radiologie-Assistenzärzte als auch Brustspezialisten eine bessere ABUS-Leistung.

Keywords

Breast - Multiview convolutional neural network - Automated breast ultrasound - Diagnostic performance - Ultrasound features

Publication History

Received: 26 September 2024

Accepted after revision: 26 June 2025

Accepted Manuscript online:
26 June 2025

Article published online:
19 August 2025

• References

• 1 Siegel RL, Miller KD, Fuchs HE. et al. Cancer statistics, 2022. CA Cancer J Clin 2022; 72: 7-33
• 2 Scheel JR, Lee JM, Sprague BL. et al. Screening ultrasound as an adjunct to mammography in women with mammographically dense breasts. Am J Obstet Gynecol 2015; 212: 9-17
• 3 Ikedo Y, Fukuoka D, Hara T. et al. Development of a fully automatic scheme for detection of masses in whole breast ultrasound images. Med Phys 2007; 34: 4378-4388
• 4 Kim SH, Kim HH, Moon WK. Automated breast ultrasound screening for dense breasts. Korean J Radiol 2020; 21: 15-24
• 5 Han S, Kang HK, Jeong JY. et al. A deep learning framework for supporting the classification of breast lesions in ultrasound images. Phys Med Biol 2017; 62: 7714-7728
• 6 Russakovsky O, Deng J, Su H. et al. ImageNet large scale visual recognition challenge. Int J Comput Vis 2015; 115: 211-252
• 7 Abdel Rahman A, Belhaouari S, Bouzerdoum A. et al. Breast mass tumor classification using deep learning. 2020 IEEE International Conference on Informatics, IoT, and Enabling Technologies (ICIoT) 2020; 271-276
• 8 Byra M. Breast mass classification with transfer learning based on scaling of deep representations. Biomed Signal Process Control 2021; 69: 102828
• 9 Wang Y, Choi EJ, Choi Y. et al. Breast cancer classification in automated breast ultrasound using multiview convolutional neural network with transfer learning. Ultrasound Med Biol 2020; 46: 1119-1132
• 10 Luo X, Xu M, Tang G. et al. The lesion detection efficacy of deep learning on automatic breast ultrasound and factors affecting its efficacy: a pilot study. Br J Radiol 2022; 95: 20210438
• 11 Lee SE, Han K, Youk JH. et al. Differing benefits of artificial intelligence-based computer-aided diagnosis for breast US according to workflow and experience level. Ultrasonography 2022; 41: 718-727
• 12 Szegedy C, Vanhoucke V, Ioffe S. et al. Rethinking the Inception Architecture for Computer Vision. In: Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition 2016; 2818-2826
• 13 Glorot X, Bengio Y. Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the Thirteenth International Conference on Articial Intelligence and Statistics. 2010; 249–256.
• 14 Mendelson E, Böhn-Véllez M, Berg WA. et al. ACR BI-RADS ultrsound. In: ACR BI-RADS atlas, Breast Imaging Reporting and Data System. Reston: Am Coll Radiol, 2013; 149.
• 15 Lee SH, Ryu HS, Jang MJ. et al. Glandular tissue component and breast cancer risk in mammographically dense breasts at screening breast US. Radiology 2021; 301: 57-65
• 16 Amin MB, Greene FL, Edge SB. et al. The Eighth Edition AJCC Cancer Staging Manual: Continuing to build a bridge from a population-based to a more “personalized” approach to cancer staging. CA Cancer J Clin 2017; 67: 93-99
• 17 Siegel S, Castellan N. Nonparametric statistics for the behavioral sciences. 2nd ed. New York: McGraw Hill; 1988
• 18 Landis JR, Koch GG. The measurement of observer agreement for categorical data. Biometrics 1977; 33: 159-174
• 19 Robin X, Turck N, Hainard A. et al. pROC: an open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinformatics 2011; 12: 77
• 20 Hillis SL, Berbaum KS, Metz CE. Recent developments in the Dorfman-Berbaum-Metz procedure for multireader ROC study analysis. Acad Radiol 2008; 15: 647-661
• 21 Lee J, Lee JH, Baik S. et al. Non-mass lesions on screening breast ultrasound. Med Ultrason 2016; 18: 446-451
• 22 Choi WJ, Kim SH, Shin HJ. et al. Automated breast US as the primary screening test for breast cancer among East Asian women aged 40–49 years: a multicenter prospective study. Eur Radiol 2021; 31: 7771-7782
• 23 Sadigh G, Carlos RC, Neal CH. et al. Impact of breast mass size on accuracy of ultrasound elastography vs. conventional B-mode ultrasound: a meta-analysis of individual participants. Eur Radiol 2013; 23: 1006-1014
• 24 Shen Y, Shamout FE, Oliver JR. et al. Artificial intelligence system reduces false-positive findings in the interpretation of breast ultrasound exams. Nat Commun 2021; 12 (01) 5645