Invasive Breast Cancers Missed by AI Screening of Mammograms: Clinical Implications and Diagnostic Challenges

on

 Invasive Breast Cancers Missed by AI Screening of Mammograms: Clinical Implications and Diagnostic Challenges

Cánceres de mama invasivos no detectados mediante mamografías con detección artificial: implicaciones clínicas y desafíos diagnósticos

AI 기반 유방촬영술 검진에서 놓친 침습성 유방암: 임상적 의미와 진단 과제

doi: 10.1148/radiol.242408

Introduction

Mammography remains the gold standard in breast cancer screening, credited with reducing breast cancer mortality by up to 50%. However, its sensitivity ranges between 75% and 85%, and drops further to 30% to 50% in women with dense breast tissue. To address these limitations, artificial intelligence (AI)-based systems have been introduced, with some outperforming radiologists in diagnostic accuracy. Despite these advances, AI algorithms are not infallible and may fail to detect certain types of breast cancers. This column provides an in-depth analysis of the false-negative rate (FNR) and the imaging features of invasive breast cancers missed by AI, based on the 2025 Radiology journal study by Woo et al.

Materials and Methods

This retrospective study reviewed mammographic data from 1082 women with 1097 invasive breast cancers diagnosed between January 2014 and December 2020. AI screening was conducted using Lunit INSIGHT MMG software. Cancers were classified based on AI abnormality score (AS), molecular subtype, imaging characteristics, and location.

AI-missed cancers were defined as lesions not detected by AI (AS <10 or incorrect location). The FNR was calculated across molecular subtypes, and the causes of misses were categorized by blinded and unblinded radiologist reviews.

Key Results

1. False-Negative Rate by Molecular Subtype

  • Luminal subtype: FNR = 17.2% (106 of 616)

  • HER2-enriched subtype: FNR = 9.0% (36 of 398)

  • Triple-negative subtype: FNR = 14.5% (12 of 83)

  • AI demonstrated the highest detection for HER2-enriched tumors due to high abnormality scores and microcalcification visibility (P < .001).

2. Characteristics of AI-missed Cancers

  • Younger age (mean 49.7 years vs 55.1 years in detected cases, P < .001)

  • Smaller tumor size (81.8% ≤ 2 cm, P < .001)

  • More luminal subtype cancers (68.8% vs 54.1%, P = .001)

  • Lower histologic grade (80.1% grade 1 or 2, P = .007)

  • More dense breasts (85.1% vs 69.0%, P < .001)

  • More BI-RADS category 4 lesions, fewer category 5

  • Higher prevalence in non-mammary zone locations (16.2%)

3. Reasons for Misses (n = 154)

  • Dense breast tissue: 72.1% of misses

  • Nonmammary zone locations (e.g., retromammary fat, subareolar area): 16.2%

  • Architectural distortion: 8.4%

  • Amorphous microcalcifications: 3.2%

4. Actionability

  • 61.7% (95 of 154) of missed lesions were deemed actionable by radiologists.

  • Actionable lesions were mostly hidden in dense breast tissue (59%), located in non-mammary zones (23%), or presented with subtle distortion or microcalcifications.

Clinical Case Examples with Figures

Figure 1: Flowchart of study inclusion and exclusion criteria.

Figure 2: A 42-year-old woman with dense breasts had an invasive ductal carcinoma in the left upper outer quadrant missed by AI due to a low AS. Confirmed post-surgery (luminal subtype).

Figure 3: A 60-year-old woman with a triple-negative 1-cm mass in the left upper inner quadrant missed by AI because of its location in the retromammary fat layer.

Figure 4: A subareolar mass in a 59-year-old woman was missed due to its location outside the mammary zone.

Figure 5: A far upper breast mass near the axilla missed by AI; histologically confirmed as grade 1 luminal carcinoma.

Figure 6: Architectural distortion in a 54-year-old woman was overlooked due to AI's failure to identify the subtle distortion.


Discussion

AI systems offer improved diagnostic efficiency and consistency, but current limitations necessitate vigilance. Key takeaways include:

  • Dense breasts remain a challenge for both human and AI interpreters. Technological enhancements to AI algorithms are required to address tissue overlap and obscured lesions.

  • Nonmammary zone locations are easily overlooked. Radiologists must maintain high suspicion in retromammary and subcutaneous areas.

  • Small, low-grade, luminal-type cancers are more likely to be missed. Enhancing AI training datasets with annotated subtle lesions may help.

  • Actionable misses highlight the continued need for radiologist oversight even with AI assistance.

Conclusion

AI mammography screening, while promising, is not foolproof. Radiologists must remain engaged in interpretation, particularly in patients with dense breasts or lesions in unusual locations. Future AI systems should be optimized to reduce FNR by incorporating more diverse training datasets, improving sensitivity for subtle findings, and enhancing detection algorithms for luminal subtype tumors.

Quiz Questions

Q1. What subtype of breast cancer had the lowest false-negative rate in AI screening?

  • (1) Luminal

  • (2) HER2-enriched

  • (3) Triple-negative

  • (4) All had the same FNR

Answer: (2) HER2-enriched
Explanation: HER2-enriched tumors had a 9% FNR, the lowest among all subtypes, attributed to higher visibility via microcalcifications.

Q2. Which factor was most commonly responsible for missed lesions by AI?

  • (1) Low histologic grade

  • (2) Microcalcifications

  • (3) Dense breast tissue

  • (4) Large tumor size

Answer: (3) Dense breast tissue
Explanation: Over 70% of missed cases were due to obscured lesions in dense breast parenchyma.

Q3. Which of the following is NOT a common location for AI-missed cancers?

  • (1) Retromammary fat layer

  • (2) Subareolar area

  • (3) Inframammary fold

  • (4) Far upper breast near axilla

Answer: (3) Inframammary fold
Explanation: The study identified other nonmammary zones but did not report the inframammary fold as a frequent miss site.

References

[1] Independent UK Panel on Breast Cancer Screening, "The benefits and harms of breast cancer screening: an independent review," Lancet, vol. 380, no. 9855, pp. 1778-1786, 2012.
[2] P. C. Gøtzsche and K. J. Jørgensen, "Screening for breast cancer with mammography," Cochrane Database Syst Rev, vol. 2013, no. 6, CD001877, 2013.
[3] L. Tabár et al., "Insights from the breast cancer screening trials," Breast J, vol. 21, no. 1, pp. 13-20, 2015.
[4] M. T. Mandelson et al., "Breast density as a predictor of mammographic detection," J Natl Cancer Inst, vol. 92, no. 13, pp. 1081–1087, 2000.
[5] E. D. Pisano et al., "Diagnostic performance of digital versus film mammography," N Engl J Med, vol. 353, no. 17, pp. 1773–1783, 2005.
[6] H. E. Kim et al., "Changes in cancer detection and false-positive recall in mammography using artificial intelligence," Lancet Digit Health, vol. 2, no. 3, pp. e138–e148, 2020.
[7] S. E. Woo et al., "Invasive breast cancers missed by AI screening of mammograms," Radiology, vol. 315, no. 3, e242408, 2025.

Comments

Post a Comment