Artificial intelligence–based computer-aided diagnosis for breast cancer detection on digital mammography in Hong Kong

Hong Kong Med J 2024;30:Epub 19 Dec 2024

ORIGINAL ARTICLE

SM Yu, MB, BS, FHKAM (Radiology)¹; Catherine YM Young, MB, BS, FRCR¹; YH Chan, MB, ChB, FRCR¹; YS Chan, MB, ChB, FHKAM (Radiology)¹; Carita Tsoi, MB, ChB, FHKAM (Radiology)¹; Melinda NY Choi, MHSc, GCB¹; TH Chan, BSc, MSc¹; Jason Leung, MSc²; Winnie CW Chu, MB, ChB, FHKAM (Radiology)¹; Esther HY Hung, MB, ChB, FHKAM (Radiology)¹; Helen HL Chau, MB, ChB, FHKAM (Radiology)

¹ Department of Imaging and Interventional Radiology, Prince of Wales Hospital, Hong Kong SAR, China

² The Jockey Club Centre for Osteoporosis Care and Control, The Chinese University of Hong Kong, Hong Kong SAR, China

Corresponding author: Dr SM Yu (ysm687@ha.org.hk)

Full paper in PDF

Abstract

Introduction: Research concerning artificial intelligence in breast cancer detection has primarily focused on population screening. However, Hong Kong lacks a population-based screening programme. This study aimed to evaluate the potential of artificial intelligence–based computer-assisted diagnosis (AI-CAD) program in symptomatic clinics in Hong Kong and analyse the impact of radio-pathological breast cancer phenotype on AI-CAD performance.

Methods: In total, 398 consecutive patients with 414 breast cancers were retrospectively identified from a local, prospectively maintained database managed by two tertiary referral centres between January 2020 and September 2022. The full-field digital mammography images were processed using a commercial AI-CAD algorithm. An abnormality score <30 was considered a false negative, whereas a score of ≥90 indicated a high-score tumour. Abnormality scores were analysed with respect to the clinical and radio-pathological characteristics of breast cancer, tumour-to–breast area ratio (TBAR), and tumour distance from the chest wall for cancers presenting as a mass.

Results: The median abnormality score across the 414 breast cancers was 95.6; sensitivity was 91.5% and specificity was 96.3%. High-score cancers were more often palpable, invasive, and presented as masses or architectural distortion (P<0.001). False-negative cancers were smaller, more common in dense breast tissue, and presented as asymmetrical densities (P<0.001). Large tumours with extreme TBARs and locations near the chest wall were associated with lower abnormality scores (P<0.001). Several strengths and limitations of AI-CAD were observed and discussed in detail.

Conclusion: Artificial intelligence–based computer-assisted diagnosis shows potential value as a tool for breast cancer detection in symptomatic setting, which could provide substantial benefits to patients.

New knowledge added by this study

With a threshold score of 30, a commercially available artificial intelligence–based computer-assisted diagnosis (AI-CAD) program showed high sensitivity and specificity for breast cancer detection on digital mammography in symptomatic settings, offering a valuable diagnostic adjunct.
The performance of AI-CAD varied according to the radio-pathological characteristics of breast cancer. Notably, the program demonstrated promising accuracy in detecting breast cancers that exhibit architectural distortion, which remains a diagnostic challenge.
Observed limitations of AI-CAD, such as underscoring cancers that present as large masses or exhibit nipple retraction as well as its inability to compare with previous studies, highlight concerns regarding standalone use of AI for triage in symptomatic clinics.

Implications for clinical practice or policy

Artificial intelligence–based computer-assisted diagnosis exhibits substantial potential for detecting breast cancers in symptomatic settings.
To make study findings clinically viable, larger validation studies are needed.

[Abstract in Chinese]

Introduction

Mammography is the principal modality used for breast cancer screening and detection in women worldwide.1 However, 10% to 30% of breast cancers may be undetected during mammography due to factors such as dense breast tissue, poor imaging technique, perceptual error, and subtle mammographic abnormalities.2

Conventional computer-aided diagnosis systems have been developed for more than two decades; however, large-scale studies have shown no significant benefit of such systems in enhancing radiologists’ diagnostic performance.3 4 Such systems do not facilitate differentiation between benign and malignant breast lesions, resulting in numerous false-positive results that require radiologist review, which may lead to reader fatigue and unnecessary additional investigations.

Currently, artificial intelligence–based computer-assisted diagnosis (AI-CAD) is widely implemented in mammography to improve diagnostic accuracy and reduce radiologist workload.5 6 The AI-CAD systems developed using deep-learning algorithms make independent decisions and self-learn without the need for feature engineering and computation.7 Artificial intelligence algorithms have been applied to multiple aspects of breast cancer screening, including risk stratification, triage, lesion interpretation, and patient recall.8 As of 2022, the US Food and Drug Administration has approved >15 AI tools for mammography applications, including density assessment, triage, lesion detection, and classification.9 Most commercial AI-CAD programs provide heatmaps with abnormality scores. Generally, higher abnormality scores indicate more suspicious radiological features and a greater likelihood of cancer.

Most existing evidence in the literature is derived from population-based screening studies.5 10 11 However, unlike other developed Asian countries such as Singapore and Korea, Hong Kong lacks a large-scale population screening programme.12 13 Our patient population primarily consists of symptomatic individuals. Evidence concerning the application of AI-CAD in symptomatic breast imaging is limited. This study aimed to evaluate the potential of AI-CAD in Hong Kong, focusing on the impact of radio-pathological phenotypes of breast cancer on AI-CAD performance. We analysed the distinctive characteristics of high-score versus low-score breast cancers. We also discuss observed strengths and limitations of AI-CAD in identifying breast cancer.

Methods

Study population

In total, 488 consecutive patients with histology-confirmed breast cancers were identified from a prospectively maintained database managed by two tertiary referral centres in Hong Kong during the period between January 2020 and September 2022. In our centres, all patients referred for diagnostic mammography were symptomatic, presenting with various clinical symptoms. We included patients with breast cancers confirmed by core needle biopsy under ultrasound guidance or stereotactic-guided vacuum-assisted breast biopsy performed at our centres. We excluded patients with diagnostic mammography performed at outside facilities (n=6), chest wall recurrence after mastectomy (n=14), cancers identified only in axillary nodes (n=3), tumour locations not feasible for mammography (n=3), and mammographically occult breast cancers undetectable by both reporting radiologists and AI-CAD (n=64) [Fig 1]. Finally, 398 patients with 414 breast cancers and 347 unaffected breasts were included in the study. Sixteen patients were diagnosed with bilateral breast cancer. Among the 382 patients with unilateral breast cancer, 35 had previously undergone contralateral mastectomy.

Figure 1. Inclusion and exclusion criteria

Image acquisition and analysis

Full-field digital mammography (MAMMOMAT Inspiration; Siemens, Erlangen, Germany or Selenia Dimensions; Hologic, Newark [DE], US) was performed prior to each biopsy. The included mammograms were exported and processed by a commercial AI-CAD program (INSIGHT MMG, version 1.10.2; Lunit, Seoul, South Korea), which is approved by the US Food and Drug Administration for lesion detection and classification in breast imaging.9

The AI-CAD algorithm used in the current study was developed and validated through multinational studies.14 15 This algorithm provides a heatmap that highlights mammographic abnormalities and generates a score ranging from 0 to 100 for each view (craniocaudal and mediolateral oblique views). The abnormality score is the maximum value for each breast, reflecting the likelihood of malignancy.

All mammograms were interpreted by radiologists subspecialising in breast radiology (with 4 to 20 years of experience in breast imaging). Mammography reports from the time of breast cancer diagnosis were retrieved from the radiology information system and retrospectively reviewed for breast density, dominant mammographic features of breast cancer, and any axillary lymphadenopathy. The clinical findings, pathological results, and molecular profiles of breast cancers were also recorded. Breast density was categorised from 1 to 4 using the BI-RADS (Breast Imaging Reporting and Data System) classification.16 The cancers were classified according to their dominant mammographic features as asymmetrical density, mass (with or without calcifications), calcifications alone, or architectural distortion.

For breast cancers presenting as a mass without calcifications, the tumour distances from the chest wall and the tumour-to–breast area ratio (TBAR) were measured in mammograms using the picture archiving and communication system by a radiologist with 2 years of experience in breast imaging. Tumour distance from the chest wall was defined as the shortest distance between the tumour and the pectoralis major in the mediolateral oblique view (Fig 2a). Tumours partially visible within the lower breast in the mediolateral oblique view, where the pectoralis muscle is not discernible, were assigned a chest wall distance of 0 cm. The TBAR was calculated via division of the tumour area by the breast area, as measured using the freehand region-of-interest tool (Fig 2b).

Figure 2. (a) Index cancer (white arrows) and measurement of tumour distance from the chest wall on the picture archiving and communication system (PACS) [double arrow]. (b) Measurement of tumour-to–breast area ratio by freehand region-of-interest on the PACS, indicated by curved arrow (tumour area) and open arrows (breast area)

The radiologists matched the index lesion to the AI-CAD heatmap to determine whether the AI-CAD correctly localised the known cancer. When the cancer was correctly localised by the AI-CAD, an abnormality score of ≥30 was regarded as a true positive, whereas a score &LT;30 was considered a false negative. When the cancer was undetected or incorrectly localised by the AI-CAD, this result also was regarded as a false negative. Breast cancers with abnormality scores of ≥90 and <30 were designated as ‘high-score tumour’ and ‘low-score tumour’, respectively.

Statistical analysis

Abnormality scores are presented as medians with interquartile ranges. The scores were analysed according to patient symptoms, breast density, mammographic findings, cancer histology, and molecular profile using the Mann-Whitney U test or Kruskal–Wallis H test. The AI-CAD abnormality scores were divided into three intervals: 0 to <30, 30-90, and >90 to 100. The Chi squared test and Mantel-Haenszel test for trend were used to analyse associations with different factors. For cancers presenting as a mass, mean abnormality scores across various TBARs and distances to the chest wall were evaluated using analysis of variance with pairwise comparisons. Statistical analyses were performed using SPSS (Windows version 26; IBM Corp, Armonk [NY], US). P values <0.05 were considered statistically significant.

Results

In total, 398 patients (mean age, 62.4 years; range, 35-100) with 414 breast cancers and 347 unaffected breasts were included in the study. The cohort consisted of two men and 396 women. Among the 414 breast cancer cases, 284 (68.6%) were palpable (Table 1).

Table 1. Median abnormality scores assigned by artificial intelligence–based computer-assisted diagnosis according to clinical, radiological, and pathological phenotypes of breast cancers (n=414)

Distribution of abnormality scores

The median and mean abnormality scores for the 414 breast cancers were 95.6 and 80.6, respectively (range, 0.4-99.9). The distribution of breast cancers according to abnormality score interval is presented in Figure 3. The sensitivity of the AI-CAD algorithm in detecting breast cancers was 91.5%, based on breast cancer identification using an abnormality score of ≥30. Overall, 65.7% of breast cancers were classified as high-score tumours, whereas 8.5% were classified as low-score tumours with abnormality scores <30; these low-score tumours were regarded as false-negative cases. Table 1 presents the medians and interquartile ranges of abnormality scores according to clinical, radiological, and pathological phenotypes.

Figure 3. Distribution of breast cancers according to abnormality score interval (n=414)

Palpable lesions, cancers in entirely fatty or scattered fibroglandular breasts, cancers presenting as masses with or without calcifications and architectural distortion, and larger cancers were associated with higher abnormality scores (all P<0.001) [Table 1]. Invasive cancers had higher abnormality scores compared with ductal carcinoma in situ (P=0.010). Axillary nodal status (P=0.078) and cancer molecular subtype (P=0.820) were not associated with abnormality scores (Table 2).

Table 2. Comparison of clinical, radiological, and pathological phenotypes of breast cancers between false-negative and truepositive results of artificial intelligence–based computer-assisted diagnosis (n=414)

Phenotypic features of high-score breast cancer

High-score breast cancers had higher prevalences of palpable disease, cancers presenting as masses with or without calcifications, invasive cancers, and larger cancers (>1 cm) [Table 2].

Phenotypic features of low-score, false-negative breast cancer

The false-negative rate for AI-CAD was 8.5% (35/414). These cancers had higher prevalences of non-palpable disease, cancers presenting as asymmetrical densities, small cancers (<1 cm), and locations in heterogeneously dense or extremely dense breast tissue.

Impact of tumour-to–breast area ratio and tumour distance from chest wall on abnormality score

Overall, 158 cancers presenting as masses without calcifications were included in this analysis. The mean abnormality score for cancers with a TBAR of ≥30% was significantly lower than for those with a TBAR of <30% (86.7 vs 54.4; P<0.001). Tumours bordering the chest wall (ie, distance of 0 cm from chest wall) demonstrated significantly lower abnormality scores compared with those located 1 cm and ≥2 cm away from the chest wall (mean, 65.5 vs 89.2 vs 87.2; P<0.001).

Distribution of abnormality scores for unaffected breasts

In the analysis of 347 unaffected breasts (regarded as negative findings by reporting radiologists), the median abnormality score was 0 (mean, 3.5; range, 0-81). Using a threshold score of 30, the false-positive rate was 3.7% (13/347), indicating 96.3% specificity. Most of these false positives (11/13) scored between 30 and 50; none scored >90. One case with known postoperative changes from breast conservative surgery showed stable mammographic finding for 10 years, scored 81 by AI-CAD. One case with a breast cyst scored 73, which was confirmed via fine needle aspiration cytology.

Discussion

Performance and potentials

Most AI-CAD algorithms provide heatmaps with abnormality scores ranging from 0 to 100; a higher score generally implies a greater likelihood of cancer. Previous AI-CAD studies have used various threshold scores; some set a threshold of 10 for population screening,18 19 20 whereas Weigel et al21 set a threshold of 28 for detecting malignant calcifications. However, the clinical implications of the abnormality score itself have not been clarified; a score range from 10 to 100 may be too broad for distinguishing malignancies in clinical practice. These aspects highlight the need for further validation of the appropriate reference score provided by AI-CAD algorithms. In this study, we set the threshold at 30 because, unlike population screening approaches, our patients were symptomatic individuals. A higher threshold score appears more practical in the clinical setting of symptomatic patients.

In our study, the AI-CAD algorithm detected 91.5% (379/414) of breast cancers with an abnormality score of >30; of these 379 cancers, 71.7% exhibited a high abnormality score of >90. The false-negative rate of 8.5% is comparable to previously reported rate for this AI-CAD algorithm.5

All cancers presenting as architectural distortion in our study were correctly localised by the AI-CAD, with abnormality scores >30; 87.5% of them were assigned high abnormality scores of >90 (Fig 4a and b). Unlike cancers presenting as masses or calcifications, cancers presenting as architectural distortion remain challenging for radiologists to detect and interpret.22 23 24 Wan et al25 showed that a standalone AI algorithm did not outperform radiologists; however, with AI assistance, junior radiologists demonstrated significant improvements in diagnostic accuracy for architectural distortion.

Figure 4. Cases illustrating the strengths of artificial intelligence–based computer-assisted diagnosis (AI-CAD). (a) A 52-year-old woman presenting with a right breast mass. The mediolateral-oblique mammographic view shows an architectural distortion (white arrow) in the upper right breast. (b) The AI-CAD program successfully detected this asymmetrical distortion within heterogeneously dense breast tissue, assigning a high abnormality score of 97. (c) A 55-year-old woman with a subtle asymmetrical density, identified as ductal carcinoma in situ on biopsy. The mediolateral-oblique mammographic view shows a subtle asymmetrical density (white arrow) in the upper left breast. The reporting radiologist did not detect the lesion on mammography but detected it via concurrent diagnostic ultrasound. (d) The AI-CAD program detected the subtle asymmetrical density, assigning an abnormality score of 68. (e) A 50-year-old woman with bilateral polyacrylamide gel implants presenting with a small lump in the left breast. The mediolateral-oblique mammographic view shows that the gel had been injected into various layers of the anterior chest wall (behind and within breast tissue, subcutaneous layer, and muscle). A subtle group of amorphous calcifications is visible in the upper left breast (white arrow). (f) The AI-CAD program detected these grouped calcifications in the context of breast augmentation, assigning a high abnormality score of 91

One case of breast cancer presenting as asymmetric density in heterogeneously dense breast tissue was missed by the reporting radiologist but detected by AI-CAD, which assigned an abnormality score of 68. The cancer was later identified by the radiologist via ultrasound, which is part of routine workup for symptomatic patients in our centre. Retrospective review indicated that the asymmetric density was visible on mammography (Fig 4c and d). In a study by Kim et al,26 40 of 128 mammographically occult breast cancers were correctly identified by the AI algorithm, demonstrating its added value in detecting such cancers.

The 64 cases of mammographically occult breast cancer not detected by either the AI-CAD or the radiologists were excluded from the study. Of these cases, 84.3% were found in heterogeneously dense and extremely dense breast tissue (BI-RADS 3 and 4).16 Dense breast tissue is recognised as a significant feature associated with mammographically occult and missed cancers.27 28 29 30 We suspect that mammographic signs of cancer are masked or obscured by dense breast parenchyma, thus evading detection by the AI-CAD. Conversely, both radiologists and the AI-CAD tended to more effectively detect cancers in fatty breasts.18

In our study, the AI-CAD correctly localised a small breast cancer with a high abnormality score (>90) in a patient with polyacrylamide hydrogel (PAAG)–injected augmentation mammoplasty (Fig 4e and f). The diagnosis of breast cancer after PAAG-injected augmentation mammoplasty is challenging. Lesion visualisation may be masked by the presence of polyacrylamide gel, and extravasated polyacrylamide gel may mimic a lesion on mammography, potentially delaying early cancer detection. In such cases, assessments of suspicious calcifications and parenchymal distortion within visible breast parenchyma are considered the main goals of screening mammography.31 32 The effectiveness of AI-CAD in detecting breast cancer among patients with augmentation mammaplasty remains uncertain, warranting further studies.

Detection challenges and future directions

Isolated cases of large, clearly visible lesions that evaded AI detection have been described by Lång et al33 and Choi et al.18 To our knowledge, our study is the first to investigate factors contributing to such evasion. In this study, the AI algorithm tended to underscore cancers presenting as large masses (Fig 5a and b). Cancers with a TBAR of ≥30% had significantly lower mean abnormality scores relative to those with a ratio of <30%. Tumours bordering the chest wall (0 cm distance) also showed significantly lower abnormality scores than those located away from the chest wall. The underlying cause remains unclear; however, these findings highlight concerns regarding the use of AI-CAD as a standalone tool for triaging cases in symptomatic populations. We also noted that the AI-CAD missed certain cancers with obvious findings, such as nipple retraction and diffuse dermal thickening (Fig 5c to f).

Figure 5. Cases illustrating the limitations of artificial intelligence–based computer-assisted diagnosis (AI-CAD). (a) A 48-year-old woman presenting with a left breast mass. The craniocaudal mammographic view shows a retracted left breast mostly replaced by a large, irregular, high-density mass with dermal infiltration and suspected pectoralis involvement. (b) The AI-CAD program detected the tumour but assigned it a low abnormality score of 30. (c) A 48-year-old woman presenting with a right breast mass. The mediolateral-oblique mammographic view shows an irregular mass with indistinct margins in the periareolar region of the right breast with nipple retraction (white arrow). (d) The AI-CAD program correctly localised the right breast mass but assigned a low abnormality score of 19, despite the presence of nipple retraction. (e) A 57-year-old woman presenting with a right breast mass. The mediolateral-oblique mammographic view shows a large right breast mass with diffuse skin thickening (white arrows). (f) The AI-CAD program detected the breast mass but assigned a low abnormality score of 32, despite the presence of diffuse skin thickening. (g) A 62-year-old woman—with a history of breast-conserving surgery for breast cancer— exhibited local recurrence on surveillance mammography. The previous mediolateral-oblique mammographic view shows postoperative changes and macrocalcification in the upper right breast; no suspicious lesion was identified. (h) The follow-up mediolateral-oblique mammographic view shows a newly developed small, irregular mass (white arrow) in the upper right breast adjacent to the macrocalcification; biopsy confirmed invasive carcinoma. (i) The AI-CAD program did not detect this lesion, assigning a low abnormality score of 8

Moreover, the inability of AI-CAD to compare mammograms with previous studies may hinder its effectiveness in specific scenarios, such as the detection of subtle developing symmetries and identification of early recurrence in postoperative cases (Fig 5g to i). In contrast, radiologists can compare mammograms with previous studies, improving mammogram interpretation accuracy.

Studies have shown that the diagnostic performances of AI algorithms are comparable to those of radiologists in terms of assessing screening mammograms; the use of AI to triage screening mammograms could potentially reduce radiologists’ workload.5 34 35 We identified potential limitations and weaknesses of AI-CAD in diagnosing breast cancers under certain conditions, highlighting the need for further large-scale studies to investigate clinical applications of AI-CAD in symptomatic patients.

Strengths and limitations

This study had several key strengths. To our knowledge, it is the first to evaluate AI-CAD for breast cancer detection in Hong Kong, using an AI-CAD system that had not previously been exposed to images from our centres during their product development. Additionally, all digital mammograms were obtained before biopsies, avoiding any biopsy-related changes which could potentially affect AI-CAD performance. Limitations of the study include its retrospective design and inclusion of cancer-enriched datasets, which may lead to overestimation of AI-CAD performance; the use of a single AI vendor, hindering applicability to other AI algorithms; and the lack of BI-RADS correlation. Furthermore, there was a lack of information concerning progression in unaffected breasts over an extended follow-up interval (≥2 years), which could impact the false-positive rate of the AI-CAD. An extended observation period is needed to identify potential malignancies that may have been initially missed by radiologists.

Conclusion

Unlike other developed cities or countries, Hong Kong does not have population-based screening programmes. The adoption and implementation of AI programs in Hong Kong for breast imaging remains in early stages, mainly due to ongoing debates about efficacy and a lack of sufficient local data to support widespread application. Current literature is almost entirely based on population screening data, which may not be applicable to cities without screening programmes. In our study, AI-CAD demonstrated promising accuracy in detecting breast cancers within symptomatic settings; its performance varied according to radio-pathological characteristics. To translate these research findings into practical clinical applications, further validation studies with larger sample sizes are required; these would confirm the reliability of AI-CAD systems. The development of protocols for integrating AI-CAD into existing clinical workflows, formulation of usage guidelines, and initiation of training programmes for radiologists to effectively utilise AI as a second reader are essential elements of this process. Collaborations with information technology departments and hospital management are necessary to ensure successful integration. Although further investigation is needed, this study provides encouraging evidence to support the use of AI-CAD as a breast cancer detection tool in symptomatic settings, ultimately benefitting patients.

Author contributions

Concept or design: SM Yu, MNY Choi, EHY Hung, HHL Chau.
Acquisition of data: SM Yu, MNY Choi, TH Chan, CYM Young, YH Chan, YS Chan, C Tsoi.
Analysis or interpretation of data: SM Yu, TH Chan, J Leung.
Drafting of the manuscript: SM Yu, CYM Young.
Critical revision of the manuscript for important intellectual content: SM Yu, CYM Young, WCW Chu, EHY Hung, HHL Chau.

All authors had full access to the data, contributed to the study, approved the final version for publication, and take responsibility for its accuracy and integrity.

Conflicts of interest

All authors have disclosed no conflicts of interest.

Funding/support

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Ethics approval

This research was approved by the New Territories East Cluster Research Ethics Committee/Institutional Review Board of Hospital Authority, Hong Kong (Ref No.: NTEC-2023-074). The requirement for informed patient consent was waived by the Committee due to the retrospective nature of the research.

References

1. Tabár L, Vitak B, Chen HH, Yen MF, Duffy SW, Smith RA. Beyond randomized controlled trials: organized mammographic screening substantially reduces breast carcinoma mortality. Cancer 2001;91:1724-31. Crossref

2. Majid AS, de Paredes ES, Doherty RD, Sharma NR, Salvador X. Missed breast carcinoma: pitfalls and pearls. Radiographics 2003;23:881-95. Crossref

3. Fenton JJ, Taplin SH, Carney PA, et al. Influence of computer-aided detection on performance of screening mammography. N Engl J Med 2007;356:1399-409. Crossref

4. Lehman CD, Wellman RD, Buist DS, et al. Diagnostic accuracy of digital screening mammography with and without computer-aided detection. JAMA Intern Med 2015;175:1828-37. Crossref

5. Dembrower K, Wåhlin E, Liu Y, et al. Effect of artificial intelligence–based triaging of breast cancer screening mammograms on cancer detection and radiologist workload: a retrospective simulation study. Lancet Digit Health 2020;2:e468-74. Crossref

6. Raya-Povedano JL, Romero-Martín S, Elías-Cabot E, Gubern-Mérida A, Rodríguez-Ruiz A, Álvarez-Benito M. AI-based strategies to reduce workload in breast cancer screening with mammography and tomosynthesis: a retrospective evaluation. Radiology 2021;300:57-65. Crossref

7. Erickson BJ, Korfiatis P, Kline TL, Akkus Z, Philbrick K, Weston AD. Deep learning in radiology: does one size fit all? J Am Coll Radiol 2018;15(3 Pt B):521-6. Crossref

8. Schünemann HJ, Lerda D, Quinn C, et al. Breast cancer screening and diagnosis: a synopsis of the European breast guidelines. Ann Intern Med 2020;172:46-56. Crossref

9. Bahl M. Artificial intelligence: a primer for breast imaging radiologists. J Breast Imaging 2020;2:304-14. Crossref

10. Hickman SE, Woitek R, Le EP, et al. Machine learning for workflow applications in screening mammography: systematic review and meta-analysis. Radiology 2022;302:88-104. Crossref

11. Leibig C, Brehmer M, Bunk S, Byng D, Pinker K, Umutlu L. Combining the strengths of radiologists and AI for breast cancer screening: a retrospective analysis. Lancet Digit Health 2022;4:e507-19. Crossref

12. Intelligence Unit, The Economist. Breast cancer in Asia—the challenge and response. A report from the Economist Intelligence Unit. 2016. Available from: https://www.eiuperspectives.economist.com/sites/default/files/EIU Breast Cancer in Asia_Final.pdf. Accessed 19 Nov 2017.

13. Lim YX, Lim ZL, Ho PJ, Li J. Breast cancer in Asia: incidence, mortality, early detection, mammography programs, and risk-based screening initiatives. Cancers (Basel) 2022;14:4218. Crossref

14. Salim M, Wåhlin E, Dembrower K, et al. External evaluation of 3 commercial artificial intelligence algorithms for independent assessment of screening mammograms. JAMA Oncol 2020;6:1581-8. Crossref

15. Kim HE, Kim HH, Han BK, et al. Changes in cancer detection and false-positive recall in mammography using artificial intelligence: a retrospective, multireader study. Lancet Digit Health 2020;2:e138-48. Crossref

16. D’Orsi CJ, Sickles EA, Mendelson EB, et al. ACR BIRADS® Atlas, Breast Imaging Reporting and Data System. Reston, VA: American College of Radiology; 2013.

17. McKinney SM, Sieniek M, Godbole V, et al. International evaluation of an AI system for breast cancer screening. Nature 2020;577:89-94. Crossref

18. Choi WJ, An JK, Woo JJ, Kwak HY. Comparison of diagnostic performance in mammography assessment: radiologist with reference to clinical information versus standalone artificial intelligence detection. Diagnostics (Basel) 2022;13:117. Crossref

19. Lee SE, Han K, Yoon JH, Youk JH, Kim EK. Depiction of breast cancers on digital mammograms by artificial intelligence–based computer-assisted diagnosis according to cancer characteristics. Eur Radiol 2022;32:7400-8. Crossref

20. Koch HW, Larsen M, Bartsch H, Kurz KD, Hofvind S. Artificial intelligence in BreastScreen Norway: a retrospective analysis of a cancer-enriched sample including 1254 breast cancer cases. Eur Radiol 2023;33:3735-43. Crossref

21. Weigel S, Brehl AK, Heindel W, Kerschke L. Artificial intelligence for indication of invasive assessment of calcifications in mammography screening [in English, German]. Rofo 2023;195:38-46. Crossref

22. Suleiman WI, McEntee MF, Lewis SJ, et al. In the digital era, architectural distortion remains a challenging radiological task. Clin Radiol 2016;71:e35-40. Crossref

23. Babkina TM, Gurando AV, Kozarenko TM, Gurando VR, Telniy VV, Pominchuk DV. Detection of breast cancers represented as architectural distortion: a comparison of full-field digital mammography and digital breast tomosynthesis. Wiad Lek 2021;74:1674-9. Crossref

24. Alshafeiy TI, Nguyen JV, Rochman CM, Nicholson BT, Patrie JT, Harvey JA. Outcome of architectural distortion detected only at breast tomosynthesis versus 2D mammography. Radiology 2018;288:38-46. Crossref

25. Wan Y, Tong Y, Liu Y, et al. Evaluation of the combination of artificial intelligence and radiologist assessments to interpret malignant architectural distortion on mammography. Front Oncol 2022;12:880150. Crossref

26. Kim HJ, Kim HH, Kim KH, et al. Mammographically occult breast cancers detected with AI-based diagnosis supporting software: clinical and histopathologic characteristics. Insights Imaging. 2022;13:57. Crossref

27. Lian J, Li K. A review of breast density implications and breast cancer screening. Clin Breast Cancer 2020;20:283-90. Crossref

28. Freer PE. Mammographic breast density: impact on breast cancer risk and implications for screening. Radiographics 2015;35:302-15. Crossref

29. Arora N, King TA, Jacks LM, et al. Impact of breast density on the presenting features of malignancy. Ann Surg Oncol 2010;17 Suppl 3:211-8. Crossref

30. Ma L, Fishell E, Wright B, Hanna W, Allan S, Boyd NF. Case-control study of factors associated with failure to detect breast cancer by mammography. J Natl Cancer Inst 1992;84:781-5. Crossref

31. Cheng NX, Liu LG, Hui L, Chen YL, Xu SL. Breast cancer following augmentation mammaplasty with polyacrylamide hydrogel (PAAG) injection. Aesthetic Plast Surg 2009;33:563-69. Crossref

32. Teo SY, Wang SC. Radiologic features of polyacrylamide gel mammoplasty. AJR Am J Roentgenol 2008;191:W89-95. Crossref

33. Lång K, Dustler M, Dahlblom V, Åkesson A, Andersson I, Zackrisson S. Identifying normal mammograms in a large screening population using artificial intelligence. Eur Radiol 2021;31:1687-92. Crossref

34. Rodriguez-Ruiz A, Lång K, Gubern-Merida A, et al. Stand-alone artificial intelligence for breast cancer detection in mammography: comparison with 101 radiologists. J Natl Cancer Inst 2019;111:916-22. Crossref

35. Rodríguez-Ruiz A, Krupinski E, Mordang JJ, et al. Detection of breast cancer with mammography: effect of an artificial intelligence support system. Radiology 2019;290:305-14. Crossref