The technologies

Artificial intelligence (AI) technologies are being used increasingly in tasks usually done by humans. These systems are trained using large datasets, with machine learning approaches. A system is provided with training images, each with the desired output or classification (termed 'ground truth' in AI). Instead of learning how to classify new cases based on predefined rules, the system learns from the examples provided and detects patterns and features that predict the output. Because the AI technology may identify features that humans do not, algorithms trained in this way can outperform humans in some classification tasks (The Royal Society, 2017).

AI technologies exist for both full field digital mammography (FFDM, 2D imaging) and digital breast tomosynthesis (DBT, 3D imaging). The AI software detects and displays suspicious features in the image, and predicts the likelihood of malignancy, to help clinical diagnosis.

This briefing focuses on 5 AI technologies for mammography: Transpara Mammography and Transpara DBT (ScreenPoint Medical), HealthMammo Software (Zebra Medical Vision), and ProFound AI for 2D Mammography and ProFound AI for DBT (iCAD). Other relevant technologies may be available but are not included in this briefing. Reasons for this include not being identified in horizon scanning because at the time, they were not commercially available to the NHS, or the company choosing not to take part.

Transpara (ScreenPoint Medical)

  • Exam type: FFDM (Transpara Mammography) or DBT (Transpara DBT).

  • Setting: screening and diagnostic.

  • Function: detecting and characterising suspicious features.

  • Algorithm: deep learning convolutional neural networks, feature classifiers and image analysis algorithms.

  • Training and validation set: over 1 million images from US and EU sites, including the OPTIMAM (NHS Breast Screening Programme) database.

  • Standalone test sets: 5,327 cases (FFDM) and 2,319 (DBT) cases representing a screening patient population.

  • Current software version: 1.6.0 (November 2019).

  • Compatible FFDM modality vendors: Fujifilm; General Electric (GE); Hologic; Philips; Siemens.

  • Compatible DBT modality vendors: Hologic; Siemens.

  • Description: Transpara is a 2-step approach to the assessment of mammograms. An overall Exam Score gives the likelihood of cancer being present in an exam. This can be used under the supervision of a radiologist or other qualified person interpreting the mammogram, to triage screening exams, or as an independent reader. Detection Aid and Region Analysis give region-based information about abnormalities found in the case, including the location, likelihood of malignancy, and type of abnormality, which can be used to support the decision of the reader.

HealthMammo Software (Zebra Medical Vision)

  • Exam type: FFDM.

  • Setting: screening.

  • Function: detecting and characterising suspicious features.

  • Algorithm: deep learning convolutional neural networks.

  • Training and validation set: more than 500,000 cases from 150 facilities across 3 continents.

  • Standalone test sets: 835 cases from the US, UK and Israel, representing a screening population.

  • Current software version: 2.2 (April 2020).

  • Compatible FFDM modality vendor: Hologic.

  • Description: HealthMammo Software automatically analyses images for suspicious findings and notifies the workstation or Picture Archive and Communication System (PACS). These findings can be used for decision support by the radiologist or other qualified person interpreting the mammogram, for worklist prioritisation and changing workflow, and as an independent second reader.

ProFound AI (iCAD)

  • Exam type: FFDM (ProFound AI for 2D Mammography) or DBT (ProFound AI for DBT).

  • Setting: screening and diagnostic.

  • Function: detecting and characterising suspicious features.

  • Algorithm: deep learning convolutional neural networks, feature classifiers and image analysis algorithms.

  • Training and validation set: approximately 2 million images.

  • Standalone test sets: 2,449 FFDM cases and 6,890 DBT cases representing a screening patient population.

  • Current software version: 2.1.

  • Compatible FFDM modality vendors: FujiFilm; GE; Hologic; Siemens, Philips.

  • Compatible DBT modality vendors: GE, Hologic, Siemens.

  • Description: ProFound AI automatically detects malignant soft tissue densities and calcifications and can be used for clinical decision support by the radiologist or other qualified person interpreting the mammogram. It records a Lesion Score for each suspicious feature detected, which represents the likelihood that the detection is malignant. A Case Score is also recorded, which indicates the likelihood that a malignancy is present in the case. This can be used to triage screening exams.


The NHS Breast Screening Programme (BSP) uses a system of 2 readers, and arbitration to interpret mammograms. However, it is currently facing a shortage of qualified people, especially radiologists. AI technologies in this setting could reduce workloads by replacing 1 of the 2 readers, or by performing triage according to likelihood of an image being malignant. This triage could be used to prioritise images according to the likelihood of malignancy, so that images with a higher chance of malignancy are reviewed sooner. It could also be used to automatically classify images showing a low likelihood of malignancy as normal, and remove these from the images to be reviewed.

Abnormalities detected within mammograms can include masses, microcalcifications, architectural distortions, and asymmetric density. Some of these changes may be small and difficult to interpret by eye, even for an experienced reader. Therefore, there is a risk of missing images that should be recalled for assessment (false negatives), and of recalling images for assessment that are normal (false positives). AI technologies could support decision making by those interpreting mammograms and reduce unnecessary or missing recalls.

Current care pathway

Women in England are invited for breast screening every 3 years from age 50 to 70, and after this they can self-refer. The NHS BSP is under the remit of the UK National Screening Committee, Public Health England. Mammograms taken in the NHS BSP are interpreted by 2 qualified independent readers. If these 2 outcomes are different, a third reader or group of readers will arbitrate. Mammography results are shared with the woman having screening and the GP within 2 weeks.

NICE's guideline on familial breast cancer includes using mammography for surveillance in those at greater risk of breast cancer, because of family history. Surveillance with mammography is also included in NICE's guideline on early and locally advanced breast cancer. In people who have had breast cancer, surveillance should be offered every year until they are eligible to be screened by the NHS BSP, or every year for 5 years if they are already of screening age.

The following publications have also been identified as relevant to this care pathway:

  • Clinical guidance for breast cancer screening assessment by the NHS BSP gives guidance for the assessment of suspicious regions detected on screening mammography. This may be done in a clinic where people with symptoms are also referred, and the assessment pathway could therefore apply to both screening and symptomatic populations. Assessment options include follow up with ultrasound, further mammography or tomosynthesis, biopsy of lesions, and less frequently, MRI. AI technologies could be used in this setting to detect and diagnose breast cancer from the mammography or tomosynthesis exams.

  • Quality assurance guidelines for breast cancer screening radiology, by the NHS BSP states that double reading of mammograms should be considered mandatory in centres offering digital mammography. Staff time could be saved if an AI technology could replace 1 of these human readers, or if it could help to target the focus of the human readers to suspicious regions.

  • The UK National Screening Committee's interim guidance for those wishing to incorporate AI into the National BSP details the evidence that would be needed to support a modification of the NHS BSP to incorporate AI. Evaluation focus will be on the impact of changes to the benefits and harms of those screened, which are linked to the spectrum of disease detected. This is therefore important to consider alongside test sensitivity. High specificity is also important to reduce false positive recalls for assessment, and their impact on downstream testing and treatment pathways. How the technology interacts with the radiologist should also be considered, and the test accuracy analysis of AI technologies, split by population subgroups such as age and ethnicity.

Population, setting and intended user

These technologies would be used in breast screening and assessment clinics. They may also have a role in clinical symptomatic clinics where mammograms may only be read once. AI software would be used by those qualified to interpret mammograms, including radiologists, and radiographers who have had appropriate training (radiography advanced practitioners).

Additional training in image interpretation would be needed, specific for the AI system used.


Technology costs

Transpara is available as an on-premises subscription model, or with a multi-year licence. Pricing is based on the volume and type of study carried out (2D FFDM or 3D DBT). This also includes installation, training, ongoing support and future upgrades. Typical prices range from £0.60 to £3.00 per exam. The local hardware that the virtual server is installed on must be provided, secured, and maintained by the customer.

HealthMammo Software runs on a server or virtual machine purchased and set up by the customer, either on-premises or in the cloud. Pricing includes installation, setup and configuration, training, a software licence and ongoing support and maintenance. Pricing is for unlimited scan volume, including cloud costs, and for each installation and PACS connection it costs £13,400 for the first year and £11,370 in subsequent years.

ProFound AI is available as a one-off purchase, with a multi-year licence, or as an on-premises subscription model. The one-off purchase includes a perpetual licence and hardware server, installation and user training, and costs in the range of £25,000 to £45,000. Additional licences can be purchased separately and cost in the range of £15,000 to £25,000, depending on the licence type. Licence types are FFDM only, DBT only, or a combination of FFDM and DBT. Post-warranty support and software updates typically cost 12% to 15% of the purchase price. Subscription pricing is based on the volume and type of study done (2D FFDM or 3D DBT). This includes installation, training, ongoing support and future upgrades. The hardware server is sold separately. Typical prices range from £1 to £3 per exam.

Costs of standard care

Investment in AI technologies would be alongside standard care. The cost of taking a mammogram is about £25. The total cost, including interpretation of results and an outpatient appointment, is about £170 (weighted average cost of consultant and non-consultant led appointments [WF01B to D and WF02B to D], NHS England National Cost Collection, 2018/19).

Resource consequences

Transpara is currently used in 4 NHS trusts, and is expected to be installed in further trusts in the next 3 to 6 months.

HealthMammo Software is not currently used in any NHS trusts.

ProFound AI is not currently used in any NHS trusts but is being used in a private sector hospital in the UK.

One expert commentator knew of about 3 symptomatic breast clinics using AI decision support systems in the UK.

Trusts adopting these technologies will need additional computing resources, including a dedicated server or virtual machine. Additional staff time and computational expertise will be needed for installation, and may also be needed for maintenance and ongoing support. Evidence suggests that adopting these technologies could reduce the workload of staff reading mammograms. It could also further reduce the time taken to read individual exams.

Training to use Transpara is given by the company and includes an introductory presentation, demonstration, hands-on training with a sample set of training exams, and training in the local clinical environment. This is estimated to take less than 2 hours in total, and takes place with the installation.

Training to use HealthMammo is provided by the company after installation, for radiologists and IT staff, and training materials are supplied for new users. Training usually takes about 1 hour.

Training to use ProFound AI is provided by the company at installation, and includes a presentation of the technology, and hands-on training with real clinical cases. On average, this takes 1 hour.

NHS England's report of the independent review of adult screening programmes in England (October 2019) states that it is widely agreed that IT systems for breast screening urgently need renewing. Multiple inefficiencies, opportunities for error and corresponding benefits that will accrue from a new system have been identified, including the prompt and economical introduction of AI software in the future.