Breast cancer diagnosis is uniquely complex and poses many challenges for pathologists. In recent years, clinical-grade artificial intelligence (AI) such as the Paige Breast Suite has been introduced to support pathologists in overcoming these challenges, as well as offer enhanced efficiency and confidence. At this year’s Digital Pathology & AI Congress: Europe, Dr. Juan Retamero, Paige’s Medical Director, Digital Pathology Transformation, and Dr. Wen Ng, Consultant Pathologist in Breast, Urology and Endocrine at St Thomas Hospital, London, hosted a session examining the value breast AI can offer labs, as well as sharing a first look at new research to continue to transform the breast cancer diagnostic experience.
Dr. Retamero began the session with a quick background on how AI is trained to be clinical-grade. He explained that Multiple Instance Learning (MIL), a weakly supervised form of training, is the best way to train models on extremely large data sets. This is critical for not only ensuring that they can distinguish between cancer and non-cancer accurately, but also making them generalizable across pre-analytical variations, which ensures that the models can then work correctly at any lab without additional tuning. For this reason, MIL is the mainstay method Paige used to train each of the AI applications in our Paige Breast Suite, a set of AI tools designed to enhance pathologists’ efficiency and confidence in diagnosing breast cancer, especially during some of the more challenging aspects of diagnosis.
For example, he shared, one challenge at the start of the diagnostic process is effectively identifying cancer in biopsy slides. In a 2015 study, it was found that concordance among pathologists in diagnosing breast cancer biopsies was only 75%, leaving room for false positives or negatives to occur.1 AI such as Paige Breast, however, has been trained to be incredibly sensitive, with an NPV of 95%, which can help pathologists more confidently classify slides as benign and reduce false positives, as well as ensure no cancers are missed.
Next, he explained that mitotic counting is not only a similarly subjective step, but a tedious and time-consuming one. AI can help point pathologists to hotspots with a high concentration of mitoses and give an accurate mitotic count to support greater efficiency and reduce grading subjectivity. AI can also offer these benefits to lymph node assessment, which Dr. Retamero explains is one of the most challenging steps. Often, the pathologists reviewing breast lymph node cases are generalist pathologists, and as one 2012 study showed, when their diagnoses were later reviewed by specialist pathologists, nearly a quarter had their N status upgraded.2 With AI like Paige Breast Lymph Node, pathologists could more confidently identify metastases, including ITCs and small challenging micromets. In fact, one study found that pathologists who utilized Paige Breast Lymph Node improved their sensitivity by 12% and reduced their overall read times by 55%.3
Of course, Dr. Retamero reminded the audience that AI, just like any other adjunct, must always be contextualized by the pathologist and treated as a tool for support, not as the absolute truth. AI is not meant to replace the pathologist, but only to save them time and improve their confidence in diagnosis.
With that, Dr. Ng took to the stage to share his own experience of applying AI into lymph node assessment. Though Dr. Ng was initially slightly trepidatious about using the technology, after he and his team conducted a pilot study with Paige Breast Lymph Node, he found that not only was the user experience incredibly simple, but there were many immediate benefits AI could offer his practice.
To start, he noted that as numbers of residents and trainees are shrinking, pathologists are being asked to complete more cases with fewer resources. At the same time, lymph node diagnosis typically takes 3-4 times more time than initial breast biopsy diagnosis and is rather mundane, yet incredibly clinically important, as it informs the prognosis and patient treatment. Paige Breast Lymph Node, he said, could dramatically reduce the time he spends on each case by automatically classifying suspicious vs non-suspicious tissue. Dr. Ng would then be able to very quickly confirm or reject the AI’s outputs and move through his caseload faster. Further, the AI’s ability to identify cancers on the case level means that he’d be able to better prioritize his worklist, reviewing suspicious cases first. Additionally, he noted that metastases can be very small and vary widely. With AI he would able to digitally annotate and measure in what he believes is a far more accurate way than with a traditional approach, and therefore could feel more confident in his overall diagnosis on even small metastases.
Importantly, all of these benefits were supported by the Pilot study. When Paige Breast Lymph Node was applied to a cohort of 53 lymph nodes from 12 patients at Dr.Ng’s lab, it offered high sensitivity and specificity for detecting cancer, and supported high intra-observer concordance. Now, he and his team are embarking on a larger version of that original study where they will review up to 1,000 lymph nodes. Though the study is in its very early stages, Dr. Ng predicts that it will demonstrate AI’s ability to provide pathologists efficiency gains. He remarked that “it is early in the study, but I can tell you the quality-of-life improvement is real. I can see already see that it’s really tremendous having that confidence that I can detect something that’s so small yet clinically potentially significant and validate that.”
Having seen what AI can do, Dr. Ng said he now hopes he can apply it in even more ways in the future. For example, he sees AI as ‘the perfect resident’ that can sort your cases for you before you even arrive at work, as well as pre-screen them as suspicious or non-suspicious, which saves time. He also foresees it being helpful for additional stain pre-ordering, IHC visualization and scoring, tumor and margin measurements, and more. He did note that there are still some lingering questions about how to practically apply AI in clinical settings that pathologists should consider, such as at what stage to introduce trainee to the AI, but overall, he said, he is very excited for the future of AI and for what the ongoing study will reveal about how it can support the diagnosis of breast cancers.
1Elmore, Joann G., et al. “Diagnostic concordance among pathologists interpreting breast biopsy specimens.” Jama 313.11 (2015): 1122-1132.
2Vestjens JHMJ, Pepels MJ, de Boer M, et al. Relevant impact of central pathology review on nodal classification in individual breast cancer patients. Ann Oncol. 2012;23(10):2561-2566. doi:10.1093/annonc/mds072
3Based on an investigational clinical study involving 3 pathologists and data from 148 patients.
*In the European Union and the United Kingdom, Paige Breast Lymph Node is approved for clinical use (CE-IVD and UKCA) with Leica Aperio AT2 and Aperio GT450 Scanners.
