Improving Breast Cancer Detection in Screening Mammography with Artificial Intelligence Assistance: A Multi-reader Retrospective Study

Hong Kong J Radiol 2026;29:Epub 26 February 2026

 

 

¹ Department of Diagnostic and Interventional Radiology, Kwong Wah Hospital, Hong Kong SAR, China

² Department of Diagnostic and Interventional Radiology, Princess Margaret Hospital, Hong Kong SAR, China

³ Department of Radiology and Organ Imaging, United Christian Hospital, Hong Kong SAR, China

 

  Full Article in PDF

In Hong Kong, breast cancer has been the most common malignancy among the female population since the early 1990s, with increasing incidence every year. It accounted for over a quarter (28.9%) of new cancer cases in 2023.[1] It was also the third leading cause of cancer deaths in women.[1] Fortunately, breast cancer can be curable in its early stages, with over 95% 5-year survival for patients with stage I disease.[2] Previous randomised controlled trials and meta-analyses have demonstrated the efficacy of screening mammography in detecting early-stage tumours and reducing breast cancer–related deaths.[3] [4] [5] [6]

 

Breast screening programmes have been established in multiple developed economies worldwide. In Western countries, the American Cancer Society recommends that women consider annual mammography screening starting at the age of 40 years,[7] whereas in the United Kingdom, the National Health Service offers breast screening every 3 years for women aged between 50 and 71 years.[8] In Asian countries, such as Japan,[9] South Korea[10] and Singapore,[11] breast screening programmes have been in place for over a decade. In Hong Kong, the Centre for Health Protection recommends that women in the general population aged 44 to 69 years with an average risk of breast cancer consider mammography screening every 2 years.[12] Together with increased advocacy from non-profit organisations, which have heightened disease awareness among the public, screening mammography has become more popular.[13]

 

Like most tests, the diagnostic accuracy of screening mammography is not absolute. Sensitivity and specificity in breast cancer detection range between approximately 50% to 80% and about 80% to 90%, respectively, in the literature.[14] [15] [16] [17] False-positive results lead to additional workup and the associated anxiety in patients, while false-negative results can delay treatment and worsen prognosis.[14]

 

Recent advancements in machine learning have led to the increased use of artificial intelligence (AI) in clinical radiology. Some studies, mainly conducted in Western countries, have shown promising results in employing AI-based tools to improve the diagnostic accuracy of screening mammography.[18] [19] [20] [21]

 

AI-supported software has become more accessible and commercially available. To the best of our knowledge, there are no published studies evaluating the diagnostic performance of screening mammography with AI assistance in Hong Kong. The lack of established evidence in our local population could be a hurdle for radiologists to consider AI-assisted screening mammography. The external validity of previous research poses a major concern. Screening mammography tests employed in studies performed in Western countries were mainly selected from Caucasian patients.[22] Asian women, on the other hand, generally have different breast composition, with a higher prevalence of dense breasts. This can obscure abnormalities on mammograms, limiting the detection of breast cancer and reducing diagnostic accuracy.[23] [24] [25] Investigations on how AI-based tools could facilitate screening mammography using test sets derived from a local Asian population could bridge this data gap.

 

This study aimed to compare the performance of radiologists in screening mammography to detect breast cancer with and without AI assistance in the local population. Subgroup comparisons between breast radiologists and general radiologists were also performed.

 

We developed a test set, the Hong Kong Personal Performance in Mammographic Screening Scheme (HKPERFORMS), to evaluate the diagnostic accuracy of radiologists in detecting breast cancer in the local Asian population with and without AI assistance. The test set comprised mammograms retrospectively selected from Asian adult female patients aged 40 years or above who underwent breast screening in a single well-woman clinic from December 2009 to December 2023. Exclusion criteria included symptomatic patients (e.g., those with a palpable breast mass), pregnant patients, and those with a history of breast implant augmentation surgery.

 

All studies in HKPERFORMS were two-dimensional (2D) screening full-field digital mammograms with standard craniocaudal and mediolateral oblique views. There were 80 mammograms showing negative or benign findings, confirmed as stable on subsequent mammographic follow-up for at least 3 years as assessed by breast radiologists recognised by the Hong Kong College of Radiologists (HKCR). There were 36 mammograms with pathologically proven breast cancer, including invasive ductal carcinoma, invasive lobular carcinoma, and ductal carcinoma in situ. Their mammographic appearances included mass (n = 21, 58.3%), calcifications (n = 6, 16.7%), architectural distortion (n = 5, 13.9%), and asymmetry (n = 4, 11.1%). The mammograms in the test set (n = 116) included breasts of varying densities: extremely dense (13.8%), heterogeneously dense (72.4%), scattered areas of fibroglandular density (12.1%), and almost entirely fatty (1.7%) [Figure 1]. Patient information and identifiers, such as name and age, were anonymised before compiled into the HKPERFORMS test set (Figure 2).

 

Figure 1. Proportion of breast densities in mammograms of the test set (n = 116).

 

Figure 2. Development of the test set.

 

This was a single-centre study. Radiologists were recruited from an acute general hospital with subspecialty training in breast radiology accredited by the HKCR. They included breast radiologists and general radiologists. Breast radiologists were defined as radiologists with at least 3 months of subspecialty training recognised by the HKCR, or post-fellowship breast radiology training, and at least 500 screening mammograms read in the past year. General radiologists were defined as HKCR members or fellows actively practising in clinical radiology, but without dedicated subspecialty training in breast radiology.

 

The recruited radiologists were blinded to all patient information and identifiers in the HKPERFORMS screening mammography test set. They assessed the mammograms under standardised conditions using dedicated software (Selenia Dimensions version 1.11; Hologic, Bedford [MA], US) with diagnostic-quality monitors (Coronis Uniti MDMC 12133; Barco, Kortrijk, Belgium) in accordance with department standards. Readers documented their screening results digitally (SurveyMonkey; SurveyMonkey, San Mateo [CA], US). Data to be entered included breast density, laterality, quadrant, depth, and presence or absence of architectural distortion if an abnormality was identified. Respondents were required to classify each study as benign or suspicious for malignancy.

 

All radiologists assessed the HKPERFORMS test set twice. In the first reading, they read the screening mammograms without AI assistance. In the second reading, additional data were provided by a commercially available AI-based tool (INSIGHT MMG version 1.1.7.3; Lunit, Seoul, South Korea),[26] which automatically highlighted regions perceived as abnormal with a colour-coded heatmap indicating the degree of suspicion. A predicted probability of malignancy was also presented numerically (Figure 3). Both pre– and post–AI-processed mammograms were available during the second reading. Respondents were instructed to record their screening results after reviewing all images. They were at liberty to follow or disregard the AI-based assessment entirely. A washout period of at least 4 weeks was observed between the two readings. The orders of the screening mammograms in the test set were different and randomised across the two sittings. Respondents who did not complete either reading were excluded from the study (Figure 4).

 

Figure 3. Screening mammogram of an adult female patient in (a) craniocaudal and (b) mediolateral oblique views with application of the artificial intelligence (AI)–based tool. A right breast upper outer quadrant mass has been colour-coded red, while central inner architectural distortion has been colour-coded green, indicating AI-perceived abnormal regions with different degrees of suspicion. Predicted probabilities of malignancy are also provided numerically. Subsequent biopsies of both lesions confirmed invasive ductal carcinoma of the right breast. The numbers on the images represent predicted probabilities of malignancy.

 

Figure 4. Assessment of screening mammograms in the test set (n = 22).

 

Background information of the recruited radiologists, including prior subspecialty training in breast radiology and experience in reporting breast imaging, was collected. All responses submitted electronically were anonymised and a random computer-generated number was assigned to each radiologist. Researchers were blinded to the identity of the respondents.

 

Statistical analysis was performed using R (macOS version 4.4.1; R Core Team, Vienna, Austria).[27] Study endpoints of diagnostic accuracy included sensitivity and specificity in the mammographic detection of breast cancer. The Obuchowski–Rockette model was used to estimate and compare diagnostic accuracy.[28] A p value of < 0.05 was considered statistically significant.

 

This manuscript was prepared in accordance with the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines.

 

A total of 22 radiologists were invited to participate in this study; six respondents who did not complete the HKPERFORMS screening mammography test set were excluded, resulting in 16 radiologists completing the test set (Figure 4). Without AI assistance, the mean sensitivity and specificity for detecting breast cancer were 73.3% and 89.9%, respectively. With AI assistance, there was significant improvement in diagnostic accuracy, with the mean sensitivity and specificity increasing to 80.7% (p = 0.007) and 94.3% (p < 0.001), respectively (Figure 5 and online supplementary Table).

 

Figure 5. Dot plots and 95% confidence intervals showing (a) sensitivity and (b) specificity of all radiologists, breast radiologists, and general radiologists in screening mammography for breast cancer detection without (red circles) and with (blue squares) artificial intelligence assistance.

 

Among the respondents, nine (56.3%) were breast radiologists and seven (43.8%) were general radiologists. The experience of the breast radiologists is shown in Figure 6. Without AI assistance, the mean sensitivity of the breast radiologists (88.3%) was significantly higher than that of the general radiologists (54.0%) in identifying breast cancer (p = 0.017). There was no significant difference in the mean specificity between the two groups (breast radiologists: 87.6% vs. general radiologists: 92.9%; p = 0.051). Using the AI-based tool, there was significant improvement in the specificity of the breast radiologists (from 87.6% to 92.6%; p < 0.001) and the sensitivity of the general radiologists (from 54.0% to 66.7%; p < 0.001). No significant changes in the sensitivity of breast radiologists and the specificity of general radiologists were observed after using the AI-based tool (Figure 5 and online supplementary Table).

 

Figure 6. Proportion of breast radiologists and general radiologists included in the study (n = 16).

 

Without assistance from the AI-based tool, the diagnostic accuracy of the breast radiologists included in this study was comparable to figures reported in the literature, with both sensitivity and specificity exceeding 85%.[15] [16] [17] In contrast, general radiologists were less likely to detect breast malignancy, with a sensitivity of about 54%. Screening tests with low sensitivity lead to a higher proportion of false-negative results, potentially leading to false reassurance and missed opportunities for early diagnosis and treatment.[14] These findings highlight the importance of dedicated training in breast radiology.[29] [30] The HKCR Mammography Statement outlines the standards for radiologists involved in screening. These include a minimum of 3 months of subspecialty training in breast radiology, interpretation of at least 500 screening mammograms annually, and ongoing participation in continuing medical education and multidisciplinary meetings.[31]

 

There were significant improvements in overall sensitivity and specificity in breast cancer detection when radiologists in this study performed AI-assisted screening mammography. This echoed previous studies which demonstrated improved diagnostic accuracy in AI-assisted mammography readings.[18] [19] [20] [21] Subgroup analysis further showed that the benefits of AI assistance differed between general radiologists and breast radiologists.

 

For general radiologists, there was significant improvement in sensitivity, from approximately 50% when screening unaided to over 65% with the use of AI-based tool. A previous study also demonstrated reduced variability in screening results and increased inter-reader reliability with AI assistance.[32] This indicates that utilising AI could yield more expertise-independent results. AI could act as an extra pair of eyes. Radiologists could refer to colour-coded heatmaps generated by AI-based software after initial mammography assessment to reduce the probability of missing breast cancer.[26]

 

Among the breast radiologists, there was improvement in specificity, while sensitivity in detecting breast cancer remained similar with and without AI assistance. The crux of screening lies in striking a balance between sensitivity and specificity. Tests with high sensitivity but low specificity may lead to over-investigation, resulting in unnecessary stress and interventions for patients.[14] While the specificity of the breast radiologists in breast cancer detection was satisfactory without AI assistance, it improved from over 85% to over 90% with the use of the AI-based tool without compromising sensitivity. Increased specificity in screening mammography would reduce call-back rates, avoid unwarranted workups for patients, and decrease the workload for radiologists.[20] [33] A study by Raya-Povedano et al[34] revealed a reduction of over 70% in radiologists’ workload following the implementation of AI-based strategies. Additionally, AI tools could be helpful to prioritise screening mammograms with suspected malignancy. Such abnormal studies could be flagged for earlier reporting by radiologists, expediting subsequent workup and treatment. Furthermore, placing flagged studies at the beginning of a screening session could minimise the risk of missed breast cancers due to reader fatigue. With the burgeoning demand for screening mammography in Hong Kong, AI-based tools could potentially alleviate the stress faced by radiologists.

 

The HKPERFORMS test set was enriched with abnormal mammograms, and the proportion of cases with biopsy-proven breast cancer was not representative of routine screening practice or the general population.[1] [2] Although respondents were instructed to interpret each individual mammogram as an independent screening case, their diagnostic accuracy might have been negatively influenced by the study design. Second, test sets used in the sittings with and without AI assistance were identical. Despite a washout period of at least 4 weeks with randomisation of the image order, radiologists might have recalled the proportion of normal to abnormal cases, potentially introducing bias in the second sitting. Third, all mammograms in the test set were 2D full-field digital mammograms. In recent years, three-dimensional mammography or digital breast tomosynthesis (DBT) has become more popular, with evidence showing improved diagnostic accuracy compared with traditional 2D mammography. Studies on AI-assisted DBT have shown non-inferior or improved sensitivity and specificity in detecting breast cancer.[35] [36] Our study did not investigate DBT performance, which remains a potential direction for further research. Finally, this was a single-centre study with limited sample size. The performance and influence of AI may vary among radiologists with differing levels of experience across diverse clinical settings. Further large-scale multi-centre investigations would provide a more comprehensive assessment.

 

This multi-reader study evaluated the potential of AI to aid breast cancer detection using HKPERFORMS, an original screening mammography test set developed from a local Asian female population with a high incidence of dense breasts. The results demonstrated that diagnostic accuracy in screening mammography was improved across radiologists with varying levels of experience in breast radiology when supported by AI-based tools.