For the first time in history, one human reader has been replaced by AI in Europe’s double reading-system, with Lunit’s AI-powered mammography analysis solution.
Lunit AI can replace one of two human readers in standard European breast cancer screening.
Addresses radiologist shortage and increasing screening demand.
The world’s first Implementation at S:t Göran Hospital reduces radiologists’ workload.
In a groundbreaking development, Europe’s traditional double reading system for breast cancer screening has seen a significant transformation. Lunit’s AI-powered mammography analysis solution has replaced one human reader, ushering in a new era of efficiency and precision in breast cancer detection. This remarkable achievement holds the potential to address the shortage of radiologists and the growing demand for breast cancer screening exams.
To come to this world’s first implementation of AI as an automated independent reader in a double-blind reading setting, we collaboratively walked a path to create strong clinical evidence.
In this blog, you will learn how we achieved this together.
Part 1 : Retrospective analysis of comparing our performance for independent assessment of Screening Mammograms
To reach this historic milestone, Lunit embarked on a collaborative journey with the Karolinska Institute and the screening clinic at S:t Göran in Stockholm, Sweden. Together, they aimed to establish robust clinical evidence supporting the implementation of AI as an independent reader in a double-blind reading setting. A critical step in this process was conducting a retrospective study comparing the performance of three commercial AI products. The results clearly demonstrated Lunit’s engine as the top-performing product, making it the ideal choice for further investigation in a prospective study.<External Evaluation of 3 commercial Artificial Intelligence Algorithms for Independent Assessment of Screening Mammograms>, Mattie Salim, MD, et al, JAMA Oncology
Part 2 : Finding the optimal threshold for Screening Mammograms
Selecting the right operating point is crucial when integrating AI algorithms into diagnostic decision-making. The choice between a high cut-off threshold for fewer flagged mammograms and a lower threshold for increased sensitivity impacts downstream workload. Previous studies primarily relied on standalone-reader matching to set the AI’s operating point, overlooking its potential synergy with human radiologists. Lunit focused on sensitivity over specificity and conducted a retrospective analysis that compared double-reading performance between AI and human readers. This approach aimed to reduce workload while maintaining diagnostic accuracy.
Figure. This figure illustrates the percent of women with cancer that are flagged as suspicious (i.e., sensitivity) by each reader. The vertical offset between the light blue and dark blue bars illustrates the discordant assessments where only one of the readers correctly flagged the case as suspicious. We show the observed measures for two radiologists and the estimated measures for AI and one radiologist, where the operating point was chosen based on standalone-reader matching (alternative 1) and on combined-reader matching (alternative 2).
Reference: <Implications for downstream workload based on calibrating an artificial intelligence detection algorithm by standalone-reader or combined-reader sensitivity matching.> Karin Dembrower ,* Mattie Salim, Martin Eklund, Peter Lindholm, and Fredrik Strand
Part 3 : Prospective study of AI replacing one human reader in Europe’s double reading environment.
The prospective study was defined by Karolinska Institute and the screening clinic at S:t Göran, Stockholm, Sweden: ScreenTrustCAD. In this study, Lunit INSIGHT MMG was in use with the aim of seeing if AI could safely and effectively replace one of the two readers.
To be able to do this a threshold needed to be set. Based on the aforementioned retrospective study on finding the optimal threshold, the threshold was chosen to attain a relative positive fraction of 1.02 for AI plus a radiologist vs. two radiologists.
ScreenTrustCAD was conducted from April 1st, 2021 to June 9th, 2022, and analyzed 55,581 breast cancer screening cases in real-world clinical settings. Using Lunit INSIGHT MMG as an independent reader, the study yielded astonishing results. Lunit, in collaboration with a single radiologist, achieved a superior Cancer Detection Rate (CDR) of 4.7 per 1000, surpassing the traditional two-radiologist approach with a CDR of 4.5 per 1000. The study also demonstrated a significant reduction in recall rates (RR) with AI, both in collaboration with one radiologist and when operating independently. These findings hold the “potential to reduce medical costs and lead to healthcare reimbursement”, as noted by Dr. Fredrik Strand, an associate professor at Karolinska Institutet.
Part 4 : Actual Implementation of AI with one human reader in Europe
Based on the compelling results of the prospective study, Lunit INSIGHT MMG has been integrated into daily clinical practice at the S:t Göran breast cancer screening clinic. In this new workflow, all cases are read by one radiologist and Lunit independently. If both classify a case as normal, no further action is taken. However, if either the radiologist or Lunit, or both, classify a case as suspicious, it undergoes a consensus meeting where radiologists have access to Lunit’s lesion-level findings. Here, it is determined whether a patient requires a recall.
Preliminary estimates indicate a remarkable reduction in radiologist reading time for one reader + AI scenario, accompanied by higher cancer detection rates and lower recall rates.
AI and human readers perceive somewhat different image features as suspicious of cancer, and thus a human reader and AI provide synergism to increase the sensitivity for detecting breast cancers in mammograms.
It is exciting to see that AI is of great help in breast cancer screening. Not only in detecting more cancers but also to keep up with the increasing number of women being screened, as well as fewer re-calls, reducing unnecessary stress and anxiety for women.
The clinical evidence of ScreenTrustCAD and the real-world implementation at S:t Göran in Sweden can serve as a template for controlled implementation for many other double-read screening settings, around the world.
“The days of double reading are numbered.”
► Read the full publication: <Artificial intelligence for breast cancer detection in screening mammography in Sweden: a prospective, population-based, paired-reader, non-inferiority study> Karin Dembrower, MD et al. THE LANCET Digital Health
If you are interested in implementing Lunit INSIGHT MMG in your breast cancer screening workflow, please feel free to reach out to us at insight@Lunit.io
Together, we can make strides in improving breast cancer screening and healthcare outcomes.