Breast Care A-Z · Concept · AIRAD

AI-assisted radiology

also: AI mammography, computer-aided detection, AI-assisted breast imaging, second-reader AI

AI-assisted radiology is software that scans your breast image — most commonly a [mammogram](/glossary/mammogram/) — and flags areas worth a second look, alongside the consultant radiologist's own reading2. The radiologist sees the image and the AI's flags together, and writes the report. It is now standard in modern UK private imaging2 and is regulated as a medical device1. See [AI in your care](/about/ai-in-your-care/) for the practice's overall approach.

Quick answers

Does the AI decide if I have cancer?

No. It is a second-look layer, not a diagnosis tool12. The consultant radiologist reads every scan and writes the report; the AI's flags are information they factor in, not a verdict. The diagnosis itself is confirmed on biopsy and histology, not on imaging.

Will the AI catch something the radiologist misses?

Sometimes — that's the point of the second-look layer. Studies of current-generation AI tools show a modest improvement in cancer-detection rate when AI is used alongside a single human reader34. The evidence is mixed enough that the radiologist's read remains the primary read and the AI is an adjunct, not a replacement2.

Can I opt out?

Hard to opt out at most centres, because AI-assisted reading is part of the standard imaging workflow1. Some imaging centres can arrange a human-only second-reader pathway at additional cost. Speak to Sarah at the practice for practical options at each clinic location.

AI-Assisted Radiology -- Breastory Encyclopaedia Plate LXXVI Plate covering AI-assisted breast radiology: deep learning mammography, CAD systems, triage, double-reading replacement, and regulatory status. TECHNOLOGY · BREAST IMAGING AI PLATE LXXVI AI-assisted radiology computer-aided detection · CAD · deep learning mammography artificial intelligence systems trained on large imaging datasets to detect, triage, and report breast cancer on mammographic and tomosynthesis images FIG 01 -- AI-assisted mammography workflow: input to radiologist decision INPUT Training data: millions of annotated mammograms DEEP LEARNING CNN Feature extraction + Classification AI OUTPUT Malignancy: 0.87 (high confidence) HIGH PRIORITY Radiologist review AI confidence score + heatmap Triage: urgent reads first RADIOLOGIST Reviews AI- annotated image at workstation Final decision: human radiologist AI = support tool not replacement i -- deep learning model (CNN trained on annotated images) ii -- feature extraction (calcifications, masses, distortion) iii -- AI confidence score + heatmap overlay iv -- triage prioritisation (urgent reads first) v -- human radiologist oversight (AI = decision support, not replacement) Approved AI uses in breast imaging Triage: prioritise urgent reads Double-reading: AI as second reader in screening QA: quality assurance Audit: detecting missed cancers retrospectively Density: automated BI-RADS density scoring Risk: lifetime risk stratification (emerging) Regulation + safety MHRA regulated: Class IIb medical device in UK Non-inferiority: Reader studies vs radiologist second reader Human oversight: Mandatory -- AI does not diagnose Accountability: Radiologist retains legal responsibility FIG 02 -- Traditional screening vs AI-assisted screening Feature Traditional AI-Assisted Reading Double human read AI + single radiologist Recall rate ~5% Similar or reduced Cancer detection Standard Non-inferior to double-read Workflow Two readers per case One reader + AI Cost Higher (two radiologists) Potentially lower Speed Standard Faster triage Miss rate Human error possible Catches different errors Radiologist workload High (radiologist shortage) Reduced Regulatory status Standard MHRA approved (UK, 2022) False positives ~5% recall Similar FIG 03 -- AI applications in breast imaging Application Description Status Mammogram triage Prioritise high-risk reads Deployed (Transpara, Mia) DBT reading Slice-by-slice AI review In use Calcification detection AI flags suspicious clusters Deployed Mass detection Irregular mass flagging Deployed Density assessment Automated BI-RADS density Deployed Risk stratification Lifetime risk from mammogram Emerging Interval cancer Retrospective audit Research Biopsy guidance Real-time USS AI Emerging Report generation Structured AI report Emerging FIG 04 -- AI-assisted screening pathway 1 Screening mammogram acquired 2 AI software analyses image (milliseconds) 3 AI assigns risk score + highlights regions of 4 Radiologist reviews AI-annotated image 5 Radiologist makes final read decision 6 Recall or routine discharge FIG 05 -- AI systems and applications AI triage (Transpara / Mia) AI second reader (screening) Automated density assessment Calcification AI detection Mass detection AI AI audit (interval cancer review) FIG 06 -- Key statistics Non-inferior to double-read in MASAI trial [1] ~50% workload reduction per reader [2] ~20-40% AI cancer detection improvement in some studies [3] >15 AI mammography systems MHRA-approved [4] References 1. Lang K et al. MASAI trial (AI screening). Lancet Oncol 2023 2. McKinney SM et al. AI mammography (Google DeepMind). Nature 2020;577:89 3. Rodriguez-Ruiz A et al. AI double-reading. JNCI 2019 4. MHRA. AI as medical device guidance 2023 5. NHS England. AI in breast screening evaluation 2023 Clinically authored by Dr Fiona Tsang-Wright , FRCS (Gen Surg) GMC 4549831 · ORCID 0000-0003-4801-026X
Visual Reference · AI-assisted radiology
Open full size ↗

Common questions The questions patients ask first

Is AI better at finding breast cancer than a radiologist?
The published evidence is mixed. Some studies of current-generation AI tools show modest improvement in cancer-detection rate when AI is used alongside a single human reader3; others show no significant difference4. The current consensus in UK practice is that AI is a useful adjunct, not a replacement, and the radiologist's read remains the primary read2.
Will the AI flag every cancer?
No. AI tools have a measurable sensitivity that is below 100%, and they miss some cancers — particularly atypical, slow-growing, or contextually-subtle ones4. This is one reason the [triple-assessment](/glossary/triple-assessment/) standard exists: no single test is sufficient.
Can I opt out of AI-assisted reading?
At most UK private imaging centres, AI-assisted reading is part of the standard workflow, applied to every scan. Opting out is harder than for tools the practice uses directly. If it matters to you, ask the imaging centre directly — some centres can offer human-only second-reader pathways at additional cost. Talk to Sarah at the practice if you want practical advice on the options at each location.
Is my image sent to the AI vendor's servers?
That depends on the deployment. Some AI tools run on the imaging centre's own servers (on-premises); some run in cloud infrastructure provided by the vendor with the imaging centre's data-protection contract in place. The imaging centre's data protection officer is the right person to ask about the specifics for your scan.
Are these AI tools approved by the NHS / regulators?
Yes. The AI tools used in UK breast imaging are regulated medical devices and carry UKCA / CE marking1. Their intended use is defined narrowly — as decision-support tools used alongside a qualified radiologist12 — and they undergo post-market surveillance after deployment.
What's the difference between AI-assisted reading and tomosynthesis?
Different things. Tomosynthesis is a type of mammogram (3D mammography) — see the [tomosynthesis glossary entry](/glossary/tomosynthesis/). AI-assisted reading is software that analyses the image, whatever the imaging modality. Most modern centres use both: a tomosynthesis image is captured, then read by both the radiologist and an AI-assisted reading tool.

AI-assisted radiology is the use of trained pattern-recognition software to scan a medical image — most commonly a mammogram — and flag areas the software considers worth a second look, alongside the consultant radiologist’s own reading of the same image. The radiologist sees both the image and the AI’s flags, and makes the final call on the report. AI-assisted radiology is increasingly used at many UK private imaging centres and is regulated as a medical device.

It is not a diagnosis-by-AI tool. It is a second-look layer: the radiologist remains the one writing the report, and the AI’s output is information the radiologist factors in, not a verdict. Many UK private imaging centres now apply AI-assisted reading alongside the radiologist’s report — confirm with your specific centre whether AI is used on your scan. At Breastory the AI tools used are the imaging centres’ tools (HCA, Circle Health Group), not Breastory’s; the radiologist who reads your scan and writes the report is the imaging centre’s consultant. See AI in your care for the practice’s overall approach.

Related terms: Mammogram · Tomosynthesis · Breast ultrasound · Breast MRI · Microcalcifications · Triple assessment · AI in your care

Definition What it is

AI-assisted radiology software is a trained machine-learning model that takes a medical image as input and produces, alongside the image, a set of regions of interest — areas the model has learned to flag as potentially containing a finding. For mammography specifically, the flags might be:

  • Suspicious masses — areas where the density or shape suggests a possible tumour.
  • Microcalcifications — clusters of small calcium deposits that can indicate early ductal carcinoma in situ (DCIS) or invasive cancer. See microcalcifications.
  • Architectural distortion — subtle pulling-in of the surrounding tissue that can indicate an underlying lesion.
  • Asymmetries — areas where the two breasts differ in a way that might warrant investigation.

The radiologist sees all of this on the same workstation as the image itself, alongside the prior images for comparison if the patient has been imaged before. The AI’s flags are typically shown as overlays — coloured boxes or contours on the image — that the radiologist can see, click, or dismiss as they read.

Evolution How it has evolved

Computer-aided detection (CAD) for mammography is not new — first-generation CAD systems were FDA-cleared in the late 1990s and have been in clinical use in some form since the early 2000s. Those early systems were crude pattern matchers and the published evidence on whether they actually improved cancer detection was inconsistent.

The current generation (introduced from around 2019 onwards) is meaningfully different — built on modern deep-learning architectures, trained on much larger and more diverse image datasets, and validated against external test sets rather than the ones they were trained on3. Where first-generation CAD was sometimes considered an interruption to the radiologist’s workflow, current-generation AI tools are typically integrated more cleanly and produce more useful flags.

Studies of current-generation AI mammography tools (notably with Lunit INSIGHT MMG, Kheiron MIA, and iCAD ProFound AI) have shown:

  • A modest improvement in cancer-detection rate when AI is used as a second-reader layer alongside a single human reader, particularly for cancers that present as subtle architectural distortion or low-density masses.
  • A reduction in radiologist workload in screening contexts where AI is used to triage clearly-normal scans for single-read while suspicious scans get the standard double-read.
  • Equivalent performance to a second human reader in some study contexts — but not all study contexts, and the evidence is mixed enough that no major UK guidance (including the UK National Screening Committee, which has not recommended AI for NHSBSP double-reading) currently recommends AI as a replacement for human reading. The NIHR-backed EDITH trial (started 2025) is testing AI as a replacement reader in NHS screening over a four-year horizon.

The current UK consensus (as of 2026) is that AI-assisted reading is a useful adjunct, not a replacement. The radiologist’s read is still the primary read; the AI’s output is information that supports it.

Limitations What it doesn’t do

Several things AI-assisted radiology does not do, despite what some marketing material implies:

  • It does not produce a diagnosis. The AI flags areas; the radiologist writes the report; the diagnosis is confirmed on biopsy and histology, not on imaging. Even if the AI flags a “high suspicion” region, that is information that triggers further investigation, not a diagnosis.
  • It does not catch every cancer. Like any pattern-matching system, AI tools have a sensitivity below 100%4. Some cancers, particularly slow-growing or atypical ones, are missed by both the AI and the human radiologist. The triple-assessment standard exists precisely because no single test is sufficient. See triple assessment.
  • It does not replace the radiologist. Every UK private imaging centre using AI tools also uses a consultant radiologist. The radiologist is the one signing the report; the AI is a tool the radiologist uses.
  • It does not see your imaging history the way a radiologist does. A human radiologist comparing your current mammogram to your previous one can spot subtle changes that AI tools sometimes miss. Most current-generation AI tools work primarily on the current image rather than longitudinally across years of priors.
  • It does not understand clinical context. The AI doesn’t know that you have a strong family history, that you came in because of a new lump, or that the previous mammogram was equivocal. The radiologist factors all of that in; the AI just looks at the image.

What it means What it means in practice for you

When you have a mammogram (or ultrasound, or MRI) at any of the imaging centres the practice uses:

  1. The radiographer takes your scan — same as it has been for years.
  2. The image is processed by the imaging centre’s AI tool within seconds of capture.
  3. The consultant radiologist reads the scan at their workstation, with the AI’s flags visible alongside.
  4. The radiologist writes the report, which goes to the practice and is shared with you at your next appointment.

You are unlikely to notice anything different from a workflow without AI. The radiographer doesn’t typically discuss the AI flags with you in the moment; the radiologist’s report is what you see.

If you want to know the specifics of which AI tool was used on your scan, the imaging centre’s data protection officer or the consultant radiologist is the right person to ask. The practice can pass that question on if it’s easier — contact Sarah at [email protected].

At consultation What to discuss with Fiona

If AI-assisted reading matters to you, raise it before your imaging is booked rather than after. Useful things to ask:

  • Which imaging centre will read my scan, and what AI tool do they use? Fiona can flag the answer or route the question to the centre’s data protection officer.
  • Is a human-only second-reader pathway available? Some centres can arrange this at additional cost; others can’t. Better to know in advance.
  • How are equivocal findings handled? Whether the radiologist flags something or the AI does, the next step is usually further imaging or a biopsy — see triple assessment for the standard pathway.

Resources Further reading

Reviewed by Dr Fiona Tsang-Wright, FRCS (Gen Surg), GMC 4549831, on 1 May 2026. Next review due 1 May 2028. If you believe anything on this page is out of date, please tell us.

Sources & guidance

Every figure on this page is anchored to a published source. Tap a number in the text or below to jump to the reference.

  1. regulatory Medicines and Healthcare products Regulatory Agency (MHRA). Software and AI as a Medical Device — Change Programme. London: MHRA. 2024 https://www.gov.uk/government/publications/software-and-ai-as-a-medical-device-change-programme Cited for: UK regulatory framework for AI-as-medical-device, including post-market surveillance and the narrowed scope of intended use.
  2. guidance Royal College of Radiologists. Artificial intelligence in radiology — position statement. London: RCR. 2024 https://www.rcr.ac.uk Cited for: UK professional consensus on AI as decision-support adjunct rather than replacement for radiologist.
  3. study Lång K, Josefsson V, Larsson AM, et al. Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI). The Lancet Oncology. 2023 ;24(8):936–944 doi:10.1016/S1470-2045(23)00298-X Cited for: Randomised trial showing AI-supported reading non-inferior to standard double-reading in population breast screening.
  4. guidance National Institute for Health and Care Excellence (NICE). Artificial intelligence as a medical device — evidence reviews. London: NICE. 2024 https://www.nice.org.uk Cited for: NICE evidence review framework for AI-as-medical-device tools.