July 12, 2023

Paige’s Research Propels AI Utilization in Pathology

Digital pathology and AI have the potential to radically transform the field of pathology for the better – and while this has become more widely known over the past few years, in practice, rigorous scientific and clinical research must first prove the safety and value of such technology before full-scale adoption can be realized.

Since it was founded, Paige has been at the forefront of this effort, conducting groundbreaking research and uncovering never-before-seen insights that have set the stage for the next era of pathology. Our work first proved that computational pathology was possible, and today continues to demonstrate the immense value of AI for detecting cancers, identifying biomarkers, and enhancing routine pathology practice.

For clinicians, our research offers a level of trust in Paige products, proving that we are consistently experimenting and improving to ensure we deliver the best possible tools for diagnosis. For the industry at large, we hope that this research accelerates the adoption of digital pathology and AI, and therefore positively impacts the entire cancer care continuum.

Here’s a look at some of the ground-breaking research that has been conducted by Paige:

AI Development

Joint Breast Neoplasm Detection and Subtyping using Multi-Resolution Network Trained on Large-Scale H&E Whole Slide Images with Weak Labels1

Based on findings accumulated through years of research and development while building Paige’s clinical products, this is the first technical publication originated from the Paige team alone, an exciting feat for our hardworking technical team.

The paper aimed to identify an approach for overcoming the problem of traditional AI training methods that require time-consuming manual annotation, which can be infeasible for large scale datasets. We proposed a weakly supervised learning framework, specifically designed to help detect, segment, and subtype breast neoplasms, which has historically been challenging.

The research found that, without training with any manual pixel-level annotations, our weakly supervised subtyping network achieved F1 scores on-par with fully supervised CNNs trained with slides collected from unseen data sources. This approach successfully created an algorithm capable of detecting invasive ductal carcinoma (IDC), invasive lobular carcinoma (ILC), ductal carcinoma in situ (DCIS), lobular carcinoma in situ (LCIS), and atypical ductal hyperplasia (ADH), and atypical lobular hyperplasia (ALH), which could prove immensely valuable for pathologists in the diagnosis of breast cancers.

AI Utilization in the Clinic

The reading paradigm: How the sequence and presentation of AI results to pathologists influences endpoints and outcome2

Critical to the successful adoption of AI in the clinic is not only understanding how the AI itself is built, but how it should be utilized in the clinic.

This study reviewed previously conducted research on Paige Prostate products to assess how the ‘reading paradigm,’ or position within the workflow at which AI is introduced, might impact pathologists. We assessed 3 possible AI use cases, including second read, concurrent read, and pre-screening.

The study revealed that the greatest time-based efficiency gains would come from a screening use case (65.5% greater efficiency) or concurrent read (over 20% time savings). Additionally, pathologists would see increased accuracy when using AI as a concurrent read and second read. Ultimately however, there were attractive benefits to using AI in all 3 methods. These findings are an important benchmark for future comparisons and research that can further support the adoption of AI in the clinic.

AI Impacts on Accuracy and Efficiency

A deep learning artificial intelligence algorithm helps pathologists improve diagnostic accuracy and efficiency in the detection of lymph node metastases in breast cancer patients3

Lymph node metastasis detection is an area where AI might be especially valuable as it has been shown to be time-consuming and challenging for pathologists. Paige performed a study to assess how our AI application designed for the purpose of supporting pathologists in this task, Paige Breast Lymph Node, would impact pathologist performance.

After pathologists reviewed a sample dataset of 167 slides with and without the assistance of Paige AI, we found that Paige Breast Lymph Node helped pathologists improve their sensitivity by 12%, from 81% to 93%. Additionally, they benefitted from an overall efficiency gain of 55%

This research should give pathologists confidence that Al can help improve diagnostic sensitivity to detect metastases of any size while reducing their reading times by more than half, which would enhance overall diagnostic efficiency and confidence.

Clinical Validation of Artificial Intelligence-Augmented Pathology Diagnosis Demonstrates Significant Gains in Diagnostic Accuracy in Prostate Cancer Detection4

Paige remains the only company with an FDA approval for AI in pathology, thanks in part to our advanced training methods and robust dataset. This study was pivotal in obtaining the approval, as it demonstrated the impact of Paige Prostate Detect, our AI application for supporting the detection of prostate cancers, on pathologist performance.

A group of 16 pathologists reviewed 610 prostate needle biopsy whole-slide images with and without assistance from Paige Prostate Detect, to find that the AI could help increase pathologist sensitivity by 8%, as well as reduce cancer detection errors by 70%. Importantly, statistically significant sensitivity gains were seen among non-GU pathologists and GU pathologists, and given that non-GU pathologists diagnose the majority of prostate biopsies, it is expected that the broad use of a tool such as Paige Prostate Detect would bring benefits to a wide portion of the patient population.

AI for Biomarker Detection

Deep learning-based assessment of HER2-low expression on breast cancer H&E digital whole slide images5

A key advantage of AI is its ability to not only simplify manual tasks, but to uncover new and innovative approaches to diagnostic challenges. A prime example is HER2 detection in breast cancer, where current testing methods such as IHC have not been optimized for finding low levels of HER2 that may now be eligible for HER2 related treatments, thanks to new medications.

In this study, Paige looked to test whether AI could help detect HER2 status on H&E slides, which is a novel approach to HER2 detection. The Al model was trained with multiple instance learning for binary classification of cases as HER2 “negative” and HER2 “expressed.” The resulting Al model could distinguish true HER2-negative from HER2-low cases with an AUC of 0.91 (+/- 0.08).

This shows that AI tools can predict HER2 status in breast cancer using H&E alone, which is an important first step for potentially enhancing the way HER2 is identified and ultimately treated.

High-throughput computational assessment of clinically relevant prostate cancer genetic phenotypes using AI analysis of H&E whole-slide images6

Similarly, prostate cancer, which is the second most diagnosed cancer among men worldwide, can require the diagnosis of certain molecular alterations, typically through the use of single or multi-gene assays that are costly and slow. Paige conducted research to assess the potential for deep learning model to identify genomic features hidden in H&E whole slide images.

The model showed a strong signal for detecting AR amplification, TP53 mutation, and RB1 and PTEN deletion, with AUCs of 0.86. This could offer a rapid option for direct clinical predictions or triage to definitive molecular testing for patient care stratification.

Advancing research in the field is critical to enabling the clinical utilization of tools such as AI. At Paige, we work tirelessly to put our applications to the test and conduct ongoing research that can not only support us in bringing the best products to market, but in furthering the practice of pathology as a whole.

To stay up to date on Paige research, subscribe to our newsletter 

1Casson A, Liu S, Godrich RA, et al. Joint Breast Neoplasm Detection and Subtyping using Multi-Resolution Network Trained on Large-Scale H&E Whole Slide Images with Weak Labels. Paper presented at: Medical Imaging with Deep Learning 2023; July 10-12 2023; Nashville, TN.

2Horton MR, Parke A, Gulturk E, et al. The Reading Paradigm: How the Sequence and Presentation of AI Results to Pathologists Influences Endpoints and Outcomes. Poster presented at: European Congress on Digital Pathology 2023; June 14-17,2023; Budapest, Hungary.

3Retamero JA, Gulturk E, Sue J, et al. A Deep Learning Artificial Intelligence Algorithm Helps Pathologists Improve Diagnostic Accuracy and Efficiency in the Detection of Lymph Node Metastases in Breast Cancer Patients. Posted presented at: United States and Canadian Academy of Pathology Annual Meeting 2023; March 11-16, 2023; New Orleans, LA.

4Raciti P, Sue J, Retamero JA, et al. Clinical Validation of Artificial Intelligence–Augmented Pathology Diagnosis Demonstrates Significant Gains in Diagnostic Accuracy in Prostate Cancer Detection [published online ahead of print, 2022 Dec 20]. Arch Pathol Lab Med. 2022. doi: 10.5858/arpa.2022-0066-OA

5Oakley J, Reis-Filho JS, Klimstra D, et al. Deep learning-based assessment of HER2-low expression on breast cancer H&E digital whole slide images. Poster presented at: San Antonio Breast Cancer Symposium; December 6-10, 2022; San Antonio, TX.

6Oakley J, Goldfinger M, Millar EKA, et al. High-throughput computational assessment of clinically relevant prostate cancer genetic phenotypes using AI analysis of H&E whole-slide images. Poster presented at Pathology Visions 2022; October 16-18, 2022; Las Vegas, NV.