Impact of AI on Breast Cancer Detection Rates in Mammography by Radiologists of Varying Experience Levels in Singapore: Preliminary Comparative Study
Bibliographic record
Abstract
Background: Breast cancer remains the most common cancer among women globally. Mammography is a key diagnostic modality; however, interpretation is increasingly challenged by rising imaging volumes, a global shortage of breast radiologists, and variability in reader experience. Artificial intelligence (AI) has been proposed as a potential adjunct to address these issues, particularly in settings with high breast density, such as Asian populations. This study aimed to evaluate the impact of AI assistance on mammographic diagnostic performance among resident and consultant radiologists in Singapore. Objective: To assess whether AI assistance improves diagnostic accuracy in mammographic breast cancer detection across radiologists with varying levels of experience. Methods: A multi-reader, multi-case study was conducted at the National University Hospital, Singapore, from May to August 2023. De-identified digital mammograms from 500 women (250 with cancer and 250 normal or benign) were interpreted by 17 radiologists (4 consultants, 4 senior residents, and 9 junior residents). Each radiologist read all cases over 2 reading sessions: one without AI assistance and another with AI assistance, separated by a 1-month washout period. The AI system (FxMammo) provided heatmaps and malignancy risk scores to support decision-making. Area under the curve of the receiver operating characteristic (AUROC) was used to evaluate diagnostic performance. Results: Among the 500 cases, 250 were malignant and 250 were non-malignant. Of the malignant cases, 16%(80/500) were ductal carcinoma in situ and 84%(420/500) were invasive cancers. Among non-malignant cases, 69.2%(346/500) were normal, 17.6%(88) benign, and 3.6%(18/500) possibly benign but stable on follow-up. Masses (54.4%, 272) and calcifications (10.8%, 54/500) were the most common findings in cancer cases. A majority of both malignant (66.8%, 334/500) and non-malignant (68%, 340/500) cases had heterogeneously or extremely dense breasts (BIRADS [Breast Imaging Reporting and Data System] categories C and D). The AI model achieved an AUROC of 0.93 (95% CI 0.91-0.95), slightly higher than consultant radiologists (AUROC 0.90, 95% CI 0.89-0.92; P=.21). With AI assistance, AUROC improved among junior residents (from 0.84 to 0.86; P=.38) and senior residents (from 0.85 to 0.88; P=.13), with senior residents approaching consultant-level performance (AUROC difference 0.02; P=.051). Diagnostic gains with AI were greatest in women with dense breasts and among less experienced radiologists. AI also improved inter-reader agreement and time efficiency, particularly in benign or normal cases. Conclusions: This is the first study in Asia to evaluate AI assistance in mammography interpretation by radiologists of varying experience. AI significantly improved diagnostic performance and efficiency among residents, helping to narrow the experience-performance gap without compromising specificity. These findings suggest a role for AI in enhancing diagnostic consistency, improving workflow, and supporting training. Integration into clinical and educational settings may offer scalable benefits, though careful attention to threshold calibration, feedback loops, and real-world validation remains essential. Further studies in routine screening settings are needed to confirm generalizability and cost-effectiveness.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.009 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.000 | 0.001 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.001 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".