AI output and image quality: reducing avoidable recalls in breast screening
Abstract
Background: Artificial intelligence (AI) is increasingly used in breast cancer screening, but its performance may be influenced by image quality, potentially leading to inappropriate recalls. Within a project funded by the Italian Ministry of Health, this study evaluated the association between AI-generated scores (Lunit INSIGHT MMG) and technical errors in screening mammography, and their impact on recall rates.
Methods: A retrospective observational study was conducted on 5,000 consecutive screening mammograms. Examinations were independently read by two radiologists, and women were recalled if at least one reader indicated further assessment. AI scores were extracted for recalled cases and stratified into three groups (1–30, 30–60, >60). Images were independently reviewed by two expert breast radiologists to identify technical errors using standardized criteria (inadequate positioning, incomplete visualization, insufficient compression, artifacts). Associations were assessed using chi-square test and logistic regression.
Results: The overall recall rate was 14.3% (714/5,000). Technical errors were identified in 150/714 exams (21.0%). Error frequency was significantly higher in the 1–30 score group (39.9%) compared with 30–60 (18.2%) and >60 (6.5%) (p < 0.001). The most frequent errors were inadequate positioning (44.7%) and incomplete visualization (30.0%), mainly associated with lower scores. Logistic regression showed a significant inverse association between AI score and technical error (OR 0.72 per 10-point increase; 95% CI 0.66–0.78). After repeat imaging without technical errors, AI flagging resolved in 97.3% of cases.
Conclusions: Image quality significantly influences AI output in mammography screening. Optimizing acquisition techniques can substantially reduce avoidable recalls and improve the efficiency of population-based screening programmes.
Key messages: • Improving image quality can significantly reduce unnecessary recalls in AI-supported screening programmes.
• Integrating AI with robust quality control can optimise resource use and screening efficiency.
Related
- Does avoidable mortality validly reflect health system quality? Avoidable deaths in West-Europe 2019
- Comment on Lan et al. Reducing Diabetic Ketoacidosis in Pediatric Type 1 Diabetes: The Impact of Screening Programs and Early Disease-Modifying Treatment. J. Clin. Med. 2026, 15, 5865
- Reducing surgical site infections among paediatric patients undergoing cardiac surgery: a multi-disciplinary quality improvement initiative
- Synthesis of the Endogenous Non-Reducing End Heparan Sulfate Disaccharides for Newborn Screening and Diagnosis of Mucopolysaccharidosis types I and II
- Investigating AI's potential in risk-stratifying lung cancer screenings for extending screening intervals and reducing workload: A retrospective analysis of a tertiary hospital cohort
BibTeX
@article{nurchis2026output,
title = {{AI output and image quality: reducing avoidable recalls in breast screening}},
author = {MC Nurchis and MT Riccardi and C Savoia and S Gigli and D Marotta and V Pettinicchio and G Damiani},
journal = {The European Journal of Public Health},
year = {2026},
month = oct,
doi = {10.1093/eurpub/ckag169.1481},
url = {https://doi.org/10.1093/eurpub/ckag169.1481},
note = {CC BY-NC 4.0},
}
RIS
TY - JOUR
TI - AI output and image quality: reducing avoidable recalls in breast screening
AU - MC Nurchis
AU - MT Riccardi
AU - C Savoia
AU - S Gigli
AU - D Marotta
AU - V Pettinicchio
AU - G Damiani
PY - 2026
DA - 2026/10/01
JO - The European Journal of Public Health
DO - 10.1093/eurpub/ckag169.1481
UR - https://doi.org/10.1093/eurpub/ckag169.1481
AB - Background: Artificial intelligence (AI) is increasingly used in breast cancer screening, but its performance may be influenced by image quality, potentially leading to inappropriate recalls. Within a project funded by the Italian Ministry of Health, this study evaluated the association between AI-generated scores (Lunit INSIGHT MMG) and technical errors in screening mammography, and their impact on recall rates. Methods: A retrospective observational study was conducted on 5,000 consecutive screening mammograms. Examinations were independently read by two radiologists, and women were recalled if at least one reader indicated further assessment. AI scores were extracted for recalled cases and stratified into three groups (1–30, 30–60, >60). Images were independently reviewed by two expert breast radiologists to identify technical errors using standardized criteria (inadequate positioning, incomplete visualization, insufficient compression, artifacts). Associations were assessed using chi-square test and logistic regression. Results: The overall recall rate was 14.3% (714/5,000). Technical errors were identified in 150/714 exams (21.0%). Error frequency was significantly higher in the 1–30 score group (39.9%) compared with 30–60 (18.2%) and >60 (6.5%) (p < 0.001). The most frequent errors were inadequate positioning (44.7%) and incomplete visualization (30.0%), mainly associated with lower scores. Logistic regression showed a significant inverse association between AI score and technical error (OR 0.72 per 10-point increase; 95% CI 0.66–0.78). After repeat imaging without technical errors, AI flagging resolved in 97.3% of cases. Conclusions: Image quality significantly influences AI output in mammography screening. Optimizing acquisition techniques can substantially reduce avoidable recalls and improve the efficiency of population-based screening programmes. Key messages: • Improving image quality can significantly reduce unnecessary recalls in AI-supported screening programmes. • Integrating AI with robust quality control can optimise resource use and screening efficiency.
SN - 1101-1262
SN - 1464-360X
ER -