Clinical and technical evidence
A literature search was carried out for this briefing in accordance with NICE's interim process and methods statement. This briefing includes the most relevant or best available published evidence relating to the clinical effectiveness of the technology. Further information about how the evidence for this briefing was selected is available on request by contacting email@example.com.
Six retrospective clinical validity studies, 3 conference proceedings, 3 conference abstracts, and 1 diagnostic accuracy study are summarised in this briefing. These include 71,470 mammography exams. Of these, 2,654 are duplicated across 2 studies, and 260 are duplicated across a study and a conference abstract. One clinical validity study, 2 conference proceedings and 1 conference abstract considered a screening population. The other study datasets were enriched with exams of people with cancer to increase the prevalence above that expected in a screening population.
One further conference abstract was provided by ScreenPoint Medical for Transpara, and a further 3 conference abstracts were provided by iCAD for ProFound AI. These abstracts did not name their respective technologies and so have not been included in the evidence summary.
Key findings include comparable or improved accuracy when artificial intelligence (AI) support is used, compared with an independent human reader, and faster reading times.
The overall evidence base is at an expected level for AI technologies emerging in the NHS. However, some of the AI technologies included in the briefing have a greater evidence base than others, in terms of both number and size of studies. The main limitation across the evidence base is that the studied datasets are mostly enriched with exams of people with cancer. However, evidence is beginning to emerge from datasets more representative of a screening population. Because the primary or intended use of these technologies is in a screening setting with lower prevalence of cancer, using enriched datasets is a significant limitation. This is recognised by the UK National Screening Committee (NSC) in their interim guidance, which requires clinical validity studies of AI technologies to be done using unenriched datasets, representative of the target population for screening. The guidance also recommends using paired, or within-person comparisons of accuracy between different AI systems, against double reading plus arbitration as the comparator. This would allow assessment of the concordance between systems, as well as their sensitivity and specificity with respect to the reference standard. Although the studies included in this briefing mostly consist of paired readings, none have provided a head-to-head comparison of AI systems. Also, most comparators were ground truth, determined by biopsy or conclusive follow up. The UK NSC interim guidance itself recommends prospective randomised test accuracy studies, preferably with multiple intervention arms to incorporate different AI technologies, and provide real world evidence. Evidence in this area is also emerging, and a study comparing 3 unnamed technologies (Salim et al. 2020) has been published. However, a further challenge in this area is that AI technologies are continuously evolving, and the evidence available is usually outdated by the time it is published. Also, any head-to-head comparisons would need to be done for comparable technologies, at the same stage in their development. A key priority of the UK NSC is to also assess interval cancers happening between routine screening visits, improved specificity, and detection of different cancer subtypes. This would be considered sufficient evidence for adoption of an AI technology that detects the same disease spectrum as standard care. For AI technologies detecting a significantly different disease spectrum, evidence of longer-term harms and benefits must be provided. Also, the optimal operating point should be established, balancing sensitivity and specificity. None of the included technologies appear to currently meet these criteria, and so the trials needed to improve the evidence base would be independent evaluation with retrospective datasets, followed by prospective randomised accuracy studies. Research is also needed to assess replacing a human reader with AI in a double reading scenario, when arbitration is needed, and the effect this has on accuracy and recall rate. The outcomes reported from the included studies are also limited. The area under the receiver operating characteristic curve (AUROC) is not clinically meaningful, and depends upon the thresholds used in practice, which may differ geographically. A further limitation of the evidence base and the technologies included in this briefing, is their lack of generalisability to all full field digital mammography (FFDM) and digital breast tomosynthesis (DBT) systems. Additionally, there is a lack of published evidence evaluating the generalisability of performance of AI technologies between different FFDM and DBT systems from the same vendor. Comparisons have not been drawn between technology performance across different screening centres. The UK NSC interim guidance states a clear preference for AI technologies that can show cross-vendor compatibility. It also indicates that those that cannot show this may only be recommended for use with FFDM and DBT systems for which evidence is available.
Mia (Kheiron Medical Technologies) is an AI technology not included in this briefing. It is currently undergoing a large-scale retrospective study of algorithm generalisability with the East Midlands Radiology Consortium (EMRAD), to assess the feasibility of its adoption as an independent reader of breast cancer screening mammograms. This is part of wave 2 of the NHS test bed programme, funded by NHS England. This briefing did not identify any published evidence for the Mia system.
Retrospective clinical validity study of AI to identify normal FFDM exams in a screening population. A cohort from the Swedish Malmö Breast Tomosynthesis Screening Trial of 9,581 double-read mammography screening exams, including 68 screen-detected cancers, and 187 false positives. The AI system assigned a cancer risk score to each exam. The effect of excluding reading by radiologists, at different thresholds, was investigated.
Intervention: Transpara (version 1.4.0, ScreenPoint Medical) to identify normal (low-risk score) mammograms in a screening population.
Comparator: ground truth, defined by histology of surgical specimen or biopsy, cross referenced to regional cancer register.
If mammograms with AI risk scores of 1 or 2 were considered normal, 1,829/9,581 (19.1%; 95% confidence interval [CI] 18.3 to 19.9) of exams could be removed from screen-reading. This included 10/187 (5.3%; 95% CI 2.1 to 8.6) false positives, and no cancers. This could reduce the workload of radiologists, and costs associated with screen-reading. Of the 5,082/9,581 (53%; 95% CI 52.0 to 54.0) with low-risk scores (5 or lower), 52/187 (27.8%; 95% CI 21.4% to 34.2%) were false positives, and 7/68 (10.3%; 95% CI 3.1% to 17.5%) were exams showing cancer. The radiologists considered all but 1 of these cancers to be clearly visible.
The AI technology had been trained and validated using about 180,000 normal, and 9,000 abnormal mammograms from 4 different vendors, independent of those used in this study. The dataset was small when the proportion of screen-detected cancers is considered. Also, the study did not try to explain why the AI system assigned low-risk scores to cancers visible to the radiologists. However, the mammograms were taken from a consecutive cohort of women presenting for population screening at a single centre in Sweden, so selection bias is unlikely. Results may not be generalisable to a UK population, although the recall rate after screening was 2.7%, which is comparable and is more representative of a screening population than studies using enriched datasets. The authors acknowledge limitations around the study population, and using only 1 mammography and AI vendor combination. The company provided technical support for the study.
Retrospective study of AI as a rule-out tool for screening mammograms. Included 18,020 double-read exams from the Danish Capital Region breast cancer screening program. There were 143 screen-detected cancers, and 447 non-cancer recalls (false positives). Exams were sorted into categories from 1 to 10 (low to high chance of malignancy) to assess the number of exams, and non-cancer recalls that could be avoided by detecting normal exams before radiologist reading.
Intervention: Transpara (version 1.5, ScreenPoint Medical) as a rule-out tool.
Comparator: outcome of double reading of screening mammograms.
At a threshold of 5, 58.52% of studies were classified as normal and included 3.5% screen-detected cancers and 23.71% non-cancer recalls. Categories 1 and 2 contained 26.29% of studies, including 5.82% non-cancer recalls and 1.36% screen-detected cancers. Category 1 contained 14.58% of studies, with 2.68% non-cancer recalls, and no screen-detected cancers.
This is a conference abstract that has not had peer review and lacks detail. The authors acknowledge that the AI technology identified some exams of people with cancer as normal exams and that improvements could be made by having radiologists examine these. The total number of exams showing cancer was also limited, and the authors highlight that a larger study should be done.
One of the authors was reported in the abstract as chief executive officer of ScreenPoint Medical, and another as an employee.
Retrospective clinical validity study of AI as a rule-out tool in FFDM. It included 18,015 consecutive screening exams acquired at a single institution on devices from 2 vendors. There were 77 exams excluded, and the final dataset included 114 biopsy-proven screen-detected cancers. Each exam was independently read by 2 radiologists, and each exam with at least 1 decision to recall for assessment was reviewed at a consensus meeting for a final recall decision. AI was used to assign an AI score to each exam, representing the chance of cancer being present, and this was used to assess the impact of setting different thresholds for exams to be read by 1 radiologist instead of 2.
Intervention: Transpara (version 1.6.0, ScreenPoint Medical) as a rule-out tool to reduce radiologist screening workload.
Comparator: all images read by 2 radiologists, with consensus meeting for all recalled cases.
All AI score thresholds were evaluated and for scores of 1 to 7, the cancer detection rate was the same as if the exam had been read by 2 readers. In some cases, cancer was missed by the AI system, but the second reader would recall it. The overall recall rate would decrease from 5.35% (n=958) with double reading of all exams, to 4.79% (n=857) if exams with scores of 1 to 7 were read by a single radiologist only, and the workload for radiologists would decrease by 32.6%.
This is from conference proceedings and may not have had peer review. The cases were collected consecutively in a screening population, so selection bias is unlikely, and the sample is representative of the target population. The authors acknowledge that no follow up was available for normal exams, and so incidence of interval cancers is unknown.
The affiliation of 1 author was ScreenPoint Medical.
Retrospective clinical validity study of AI as a tool to rule-in using DBT after FFDM. An AI score of the chance of cancer being present was given by the system to 14,768 FFDM exams, which also had DBT exams available. The effect of different AI score thresholds for adding a DBT exam to FFDM, on number of cancers detected, additional DBT exams needed, detection rate, and false positives, was assessed.
Intervention: Transpara (version 1.4.0, ScreenPoint Medical) as a rule-in for DBT exam after FFDM.
Comparator: FFDM images read by Transpara (version 1.4.0, ScreenPoint Medical) only.
At an AI score threshold of 9.0, 26% (n=25) more cancers would be detected than on FFDM alone. This is 61% of the 41 cancers originally detected only on DBT, and 16 (12%) would still be missed. At this threshold, 1,797 (12%) of exams would have both FFDM and DBT.
Retrospective clinical validity study of AI as a rule-in tool. There were 2,683 screening mammograms from the OPTIMAM Medical Image Database (OMI-DB), which collects NHS Breast Screening Programme images from centres in the UK. All had biopsy-proven cancers of different grades and 1,212 had a previous mammogram available. The AI system was calibrated to make recall decisions at 50%, 10% and 4% recall rates, using a separate independent screening dataset.
Intervention: Transpara (ScreenPoint Medical) as a rule-in tool.
Comparator: biopsy-proven ground truth.
When applied to the 2,683 mammograms with biopsy-proven cancer, the calibrated AI had sensitivities of 99.3%, 87.7% and 76.1% for recall rates of 50%, 10% and 4%, respectively. At a recall rate of 4%, 16.8% of the 1,212 previous screening mammograms would have been recalled. Also, at a 4% recall rate, the AI recalled a greater proportion of higher than lower grade cancers (80.7% grade 3 compared with 68.2% grade 1) at p<0.001.
This is a conference abstract that has not had peer review and lacks detail. The mammograms used in the study had not been used previously to train, validate or test the AI system. The results may be generalised to the UK because they used UK cases, and a recall rate of 4% to assess previous screening mammograms, which is typical in a UK screening setting. However, these mammograms were selected because of a later positive mammogram, and do not represent a typical screening population.
One of authors was reported in the abstract as an employee of ScreenPoint Medical, and another as a shareholder.
Retrospective clinical validity study of AI as a stand-alone reader in FFDM assessing 2,654 digital mammography exams in study dataset. Ground truth was verified by histopathological analysis or follow up as: cancer (n=653), benign (n=768) and normal (n=1,233). About half of exams were from a screening population and half were from a clinical symptomatic population, collected from studies across 7 countries using mammography systems from 4 different vendors.
Intervention: Transpara (version 1.4.0 ScreenPoint Medical).
Comparator: average radiologist performance against ground truth.
AI performance was statistically non-inferior to that of the average of 101 radiologists. The AUROC for the AI system was 0.840 (95% CI 0.820 to 0.860) compared with 0.814 (95% CI 0.787 to 0.841). AI had AUROC higher than 61.4% of radiologists.
The AI technology had been trained and validated using over 189,000 mammograms (5% with cancer), independent of those used in this study. However, the source of these images was not reported. The mammograms used in this study were from studies across different countries, so the results are likely generalisable to a large population. It is reported that there is overlap between 2 of the studied datasets, and that centres provided a Breast Imaging Reporting and Data System BI-RADS score. This indicates how concerned the interpreting reader is about the findings, or probability of malignancy, which could cause inconsistencies in interpretation. The study datasets were enriched with malignant cases, which is not normal for the intended screening population. Knowing this may have influenced the judgement of the radiologists scoring the images. There is a discrepancy in the number of mammograms stated in the abstract and main body of the text.
Two of the 15 authors are named in the paper as employees of ScreenPoint Medical, and technical support for the study was provided by the company.
Intervention: Transpara (version 1.4.0, ScreenPoint Medical) as a rule-out tool to reduce radiologist screening workload.
Comparator: all images read by radiologist.
Setting the likelihood threshold at 5 (high likelihood above 5) resulted in a trade-off between approximately halving (-47%) the radiologist workload, and excluding 7% of true-positive exams. A threshold of 2 resulted in workload reduction of 17%, and excluding only 1% of true positives. The area under the curve was not affected by pre-selection except at the extreme AI-generated likelihood score of 9.
In addition to limitations for related study, Rodriguez-Ruiz et al. (2019a), the authors stated that they were unable to analyse results for detection mode (screening or clinical), for histopathological type of cancers, or for breast density, because this information was not available from the original studies.
Retrospective clinical validity study of AI as a decision support aid in FFDM for 240 (100 cancer, 100 normal, 40 false positive cases recalled for assessment) digital mammography exams from 2 different mammography vendors in the US and Europe. Fully crossed study design with 14 radiologists reading half the images with AI support and half without AI support, in 2 sessions at least 4 weeks apart.
Intervention: images read by radiologists with Transpara (version 1.3.0 ScreenPoint Medical) AI decision support.
Comparator: images read by same radiologists without AI support.
AUROC was higher with AI support (0.89 compared with 0.87, p=0.002), sensitivity increased with AI support (86% compared with 83%, p=0.046), specificity trended toward improvement but not significantly so (79% compared with 77%, p=0.06). Reading times per case were similar (149 seconds with AI compared with 146 seconds without AI). AUROC of the AI system alone was similar to average AUROC of the radiologists (0.89 compared with 0.87).
The sample size was small, and the method for randomly selecting the final population from the available exams of each type was not reported. The study covered the US and Europe, giving some generalisability to populations in those areas. Only exams from 2 device vendors were considered, so the AI performance on exams taken on other devices cannot be assumed. Despite the main inclusion criteria being women attending for screening with no symptoms or concerns, the selection of cases produced a dataset enriched with exams of people with cancer. This is therefore not representative of the intended population. Although the mammograms were collected in 2 centres, all radiologists were based in the US, and interpretation may differ geographically or locally.
ScreenPoint Medical was responsible for data generation, and payment of external study costs. A research agreement exists between the academic institution responsible for the work and the company. However, no financial compensation was provided.
Retrospective study of AI to replace 1 radiologist in a double reading setting. In total, 31,650 radiologist assessments of 2,892 mammograms were available in a database. Bootstrapping was used to select different combinations of mammogram and radiologist assessment. This was to simulate scenarios of double human reading and double hybrid reading with the second reader replaced by AI. For this study, an AI score of 10, indicating chance of cancer, represented a recall by the system.
Intervention: images read by radiologists with Transpara (version 1.4.0, ScreenPoint Medical) as second reader. Consensus by an independent radiologist if needed.
Comparator: images read by 2 independent radiologists, with consensus by an independent radiologist if needed.
When using AI as a second reader, the workload (number of human assessments including arbitration) was reduced by 44%, compared with double human reading. Sensitivity was similar, at 81.4% compared with 81.5% (p=0.88). Specificity was improved by 5.3% with AI from 69.9% to 75.2% (p<0.001).
This is from conference proceedings and may not have had peer review. The authors acknowledge that using a single reader to arbitrate may not be representative of centres who use panels for this. The dataset was enriched with exams of people with cancer. It also used mammograms from a range of previous studies, which have their own limitations. It is not clear if the same images and radiologist readings were used for both scenarios in each case. These mammograms are reported to have been used previously to benchmark the standalone performance of the AI system, compared with that of radiologists.
The affiliation of the lead author was ScreenPoint Medical.
Retrospective clinical validity study of AI as a standalone reader of FFDM exams for 310 people of Japanese family origin in an outpatient setting, including 69 with malignant lesions. The AI system gave each case a cancer likelihood score, and recall thresholds of 4 and 7 were used to determine sensitivity and specificity.
Intervention: Transpara (version 1.3.0, ScreenPoint Medical).
Comparator: performance of 3 radiologists reaching a consensus, against ground truth.
The AUROC was higher for human readers than for Transpara, at 0.816 compared with 0.706 (p<0.001). Sensitivities were 89% for human readers. For AI they were 93% for a cancer likelihood cut-off score of 4, and 85% for a cut-off of 7. Specificities were 86% for the human readers, and 45% and 67% for AI, with cut-offs of 4 and 7 respectively.
The AI technology had been trained and validated using a multi-vendor multicentre mammogram database. This included 9,000 biopsy-proven cancers, independent of those used in this study. The small number of exams studied were selected by a radiologist from 11,891 exams so that 22% had malignant lesions. The cases were therefore prone to selection bias, and not representative of a screening population. The people in the study were of Japanese family origin, so results may not be generalisable to a UK population. The authors acknowledge limitations around training the AI system using mammograms mainly from a Western population (Transpara's training and validation set is based on images from US and EU sites).
Retrospective clinical validity study of AI as a decision support tool in DBT using 260 DBT exams from 7 sites in the US. These included 65 exams found to have cancer. This was a fully crossed study design with 24 radiologists (13 of whom were breast subspecialists) reading 260 images with AI support and 260 without AI support, in 2 sessions at least 4 weeks apart.
Intervention: exams read by radiologists with PowerLook Tomo Detection 2.0 decision support (since rebranded as ProFound AI, iCAD).
Comparator: exams read by same radiologists without AI support.
With AI, average radiologist AUROC increased by 5.7% (95% CI 2.8 to 8.7), and reading time was improved by 52.7% (95% CI 41.8 to 61.5). Case-level sensitivity and lesion-level sensitivity increased by 8.0% (95% CI 2.6 to 13.4) and 8.4% (95% CI 2.9 to 13.9) respectively. Specificity increased by 6.9% (95% CI 3.0 to 10.8), and recall rate in non-cancers decreased by 7.2% (95% CI 3.1 to 11.2). All results were statistically significant (p<0.01).
Exams selected were enriched with exams of people with cancer, but readers were blinded to proportions. No study cases were used for development or training of algorithm. Those reading exams had not practiced at the acquisition sites. There was a randomised reading order to limit recall bias. Readers knew they were being timed but were blinded to the measurement. Sample size calculation were used, and study end points were explicitly stated. The method of random selection of exams from those eligible for analysis was not disclosed.
Study funded by iCAD, and 6 authors are either employees of or consultants for the company. Data and publication were controlled by an author not affiliated with the company.
Intervention: exams read by radiologists with ProFound AI (version 2.0, iCAD) decision support.
Comparator: exams read by same radiologists without decision support.
Outcomes reported included those reported in Conant et al. (2019a), plus outcomes relating to subgroups of people with cancer. Average radiologist sensitivity improved 6.8% with AI for soft tissue lesions, 6.2% for invasive carcinomas, 12.0% for calcifications only, and 14.6% for ductal carcinoma in situ (DCIS). Specificity improved 7.9% with AI for soft tissue, and 8.4% for lesions with BI-RADS classifications of 1 or 2. For calcifications only, average radiologist specificity reduced 2.7%, and for all exams not showing cancer it improved 6.9%. Reading times in exams showing cancer were reduced 39.5% for soft tissue, 41.4% for invasive carcinomas, 46.8% for calcifications, and 40.3% for DCIS. For exams without cancer, reading times were reduced 54.4% with AI for soft tissue lesions, 59.1% for BI-RADS classifications of 1 or 2, 43.6% for calcifications only, and 55.9% overall for all of these exams.
Diagnostic accuracy study of AI in FFDM, including a retrospective cohort of 835 FFDM screening mammograms from the UK, US and Israel. Ground truth classifications were 435 biopsy-confirmed cancers, and 400 exams with no cancer, confirmed by 2‑year follow up. Mammograms were selected to include a representative range of lesion types, breast densities, ages, and histology types.
Data is unpublished from Food and Drugs Administration section 510(k) premarket notification: performance data section.
Intervention: HealthMammo Software (Zebra Medical Vision).
Comparator: ground truth mammograms showing cancer confirmed by biopsy and mammograms not showing cancer confirmed by a negative follow up of 2 years.
HealthMammo showed an AUROC of 0.9661 (95% CI 0.9552 to 0.9769). In standard mode, the sensitivity was 89.89% (95% CI 86.69 to 92.38) and specificity was 90.75% (95% CI 87.51 to 93.21). In high sensitivity mode, sensitivity was 94.02% (95% CI 91.39 to 95.89) and specificity was 83.5% (95% CI 79.55 to 86.82). In high specificity mode, the sensitivity was 84.41% (95% CI 80.41 to 87.27) and specificity was 94.00% (95% CI 91.23 to 95.94).
The study is unpublished and so not peer reviewed. It uses a relatively small sample size. Images are from 3 countries including the UK, and so could be representative of target population. There is a risk of selection bias because it was not reported how included cases were selected. Although intended for screening, the dataset is enriched with exams of people with cancer, and no comparison has been made with human readers.
ScreenPoint Medical advised that several international clinical studies of Transpara are ongoing using unenriched datasets. The focus of these is early detection, reducing interval cancers and workload reduction. Also, 2 prospective clinical trials using Transpara are being set up. Several papers are currently under review in journals, with several more in preparation. There are 6 presentations planned for the Radiological Society of North America (RSNA) 2020 annual meeting.
Zebra Medical Vision advised that several international clinical studies are ongoing. These use HealthMammo to validate and collect evidence of the benefits of using the system in a screening setting.
iCAD advised that several international clinical studies of ProFound AI for 2D Mammography and ProFound AI for DBT are ongoing. These include a focus on reducing the number of interval cancers and improving cancer detection rates in screening.