TL;DR
A single-institution review of 414 women with breast cancer found an AI-based computer-aided detection system failed to both flag and localise 127 cancers (30.7%). Radiologists reviewing only diffusion-weighted MRI (DWI) identified most of those AI-missed lesions, suggesting DWI could be a useful adjunct, though broader validation is needed.
What happened
Researchers reviewed imaging from 414 women (mean age 55.3) recently diagnosed with breast cancer to assess an AI-based computer-aided detection (AI-CAD) tool on mammography and breast MRI. For a lesion to count as detected, the system had to mark it as suspicious and localise it correctly; lesions that did not meet both criteria were classified as AI-missed. The AI-CAD failed to detect 127 cancers, representing 30.7% of cases. Two factors were strongly associated with misses: dense breast tissue and smaller tumour size — cancers 2 cm or smaller were nearly five times more likely to be missed. To test a countermeasure, two radiologists read only the diffusion-weighted imaging (DWI) portion of MRI scans, a fast contrast-free sequence. DWI readings identified the majority of AI-missed cancers (83.5% for one reader, 79.5% for the other) with substantial interreader agreement. DWI performed best for tumours >1 cm and for lesions not seen on mammography, but its sensitivity declined for lesions under 1 cm. The authors note the study’s limits—participants all had known cancer and the work was done at a single centre—and call for prospective, multicentre trials to confirm the findings.
Why it matters
- AI tools are increasingly used in breast imaging; a high miss rate could affect diagnostic reliability.
- Dense breast tissue and small tumours remain major blind spots for the evaluated AI system.
- DWI, a rapid contrast-free MRI technique, may detect many cancers the AI misses and could act as a safety net.
- Findings highlight the need to validate AI performance across diverse screening populations before wider reliance.
Key facts
- Study sample: 414 women with recently diagnosed breast cancer; mean age 55.3 years.
- AI-CAD missed 127 cancers — 30.7% of cases — by the study’s detection criteria (suspicious score plus correct localisation).
- Small tumours (≤2 cm) were almost five times more likely to be missed by the AI system.
- Dense breast tissue was strongly associated with lesions missed by the AI.
- Two radiologists read only diffusion-weighted imaging (DWI) from MRI: detection of AI-missed lesions was 83.5% for one reader and 79.5% for the other.
- Interreader agreement on DWI interpretations was described as substantial.
- DWI performed best for tumours larger than 1 cm and for lesions occult on mammography; sensitivity fell for lesions smaller than 1 cm.
- Study limitations: included only women already diagnosed with cancer and was conducted at a single institution.
- Authors recommend prospective, multicentre trials to confirm whether DWI can reliably augment AI-assisted detection.
What to watch next
- Results of prospective, multicentre trials testing DWI as an adjunct to AI-assisted breast imaging (confirmed in the source).
- Validation studies that assess AI-CAD performance across routine screening populations, including women without a prior cancer diagnosis (confirmed in the source).
- Development of AI algorithms specifically trained to address dense breast tissue and small tumour detection (not confirmed in the source).
Quick glossary
- AI-CAD: Artificial intelligence–based computer-aided detection/diagnosis systems used to assist radiologists by identifying and scoring suspicious findings on imaging.
- Diffusion-weighted imaging (DWI): An MRI sequence that measures the movement of water molecules in tissue; it is fast and does not require contrast injection.
- Breast density: A radiologic description of the amount of fibroglandular tissue in the breast; higher density can mask tumours on mammography.
- Mammogram: An X-ray image of the breast commonly used for screening and initial evaluation of breast abnormalities.
Reader FAQ
How many cancers did the AI system miss?
The AI-CAD failed to both flag and correctly localise 127 of 414 cancers, 30.7% of cases.
Which cancers were most often missed by the AI?
Cancers in dense breasts and smaller tumours (≤2 cm) were more likely to be missed.
Did DWI detect the cancers the AI missed?
Radiologists reviewing only DWI identified most AI-missed lesions (83.5% and 79.5% for two readers), with substantial agreement between them.
Should patients rely on DWI instead of AI or mammography?
not confirmed in the source
Will these results change screening guidelines?
not confirmed in the source
This site is intended for healthcare professionals EUR USA Home Radiology AI Misses Nearly One-Third of Breast Cancers, Study Finds 2 Dec 2025 Radiology View All News Press play to…
Sources
- AI misses nearly one-third of breast cancers, study finds
- AI Catches One-Third of Interval Breast Cancers Missed at …
- Nationwide real-world implementation of AI for cancer …
- How Artificial Intelligence is Reshaping Breast Cancer …
Related posts
- MRI study finds lasting brain changes after COVID-19 infection
- Go.sum is not a lockfile — stop parsing it to analyze dependency graphs
- Lean 4 proof argues SSOT needs definition-time hooks and introspection