Digital pathology sits at the intersection of pathological diagnosis and precision medicine. In recent years, 3D pathology imaging has developed to achieve precision diagnosis, which can address some challenges in diagnosis excellence.1 Diagnosis is the foundation of effective and quality care; it has profound influence on patient outcomes.2
Using AI/ML to reduce the cognitive demands on clinicians and/or the associated diagnostic errors could help reduce adverse consequences of incorrect or delayed diagnosis in several medical specialties, such as radiology and oncology.3,4 Before leveraging AI/ML’s capabilities to support analysis and interpretation of digital images, and further aiding precision medicine by enabling more individualized pathological diagnosis, identification of suitable “biomarker ensembles” and key parameters for proper classification of patients for a particular disease (i.e., effective companion diagnostics) is needed.5 This understanding and validation will enable application of diagnosis excellence principles, such as better linking diagnosis to treatment, providing more thoughtful interpretation of tests anticipating the probability of specific diagnostic errors, and more consistently quantifying test data to communicate benefits and limitations of diagnostic results.6
In addition, the current probability of success for oncology drug development is low despite its high costs. One of the main difficulties is achieving precision in patient screening, which relies on a pathology diagnosis. There are growing demands for precision cancer diagnosis but an insufficient number of pathologists to make the diagnoses. Therefore, the digital pathology market has grown over the past 10 years. The immuno-oncology drug market size is also increasing, yet tumor microenvironment evaluation is difficult with conventional 2D pathology. Despite playing a key role in clinical decisions for patient treatments for over 150 years, conventional 2D pathology deals with thin slides containing only a small portion of a patient’s biopsy (about 100 cells in a thin slide of a needle biopsy). Formalin-fixed paraformaldehyde embedded (FFPE) tissue blocks contain abundant medical information (at μm resolution) about individual patients, yet the current pathology workflow using visible light reagents is not applicable for labeling spatial features of morphology or biomarker distribution in-depth. Therefore, valuable 3D pathology information of patients in FFPE remains yet to be fully utilized.
Recent Advancements In Digital Pathology And New AI/ML Applications
The advantages of 3D imaging over 2D can be illustrated by the comparison of computed tomography (CT) to X-ray in radiology imaging. It took quite some time to progress from digitized X-ray film to digital image to CT. One application of X-ray coupled with CT diagnosis is giving more precise parameters for evaluation of COVID-19 infected patient status.7 Digital pathology trends resemble the age of transforming X-ray film to digital images, yet development of 3D digital pathology to support precision diagnosis and precision medicine is quickly evolving. In fact, medical images of large data size have been digitized and are widely used for training AI models, among which radiology images (at mm resolution) derived from X-rays and CT images are thoroughly investigated to develop FDA-approved products.8 One example successfully implemented in hospitals to identify intracranial hemorrhage in support of better and faster clinical decisions is DeepCT, with an AI-powered triage system on head CT images developed by Deep 01 (approved by the FDA in the U.S., Taiwan, and Japan).
In the case of lung cancer and immuno-oncology therapy, a treatment modality (immune checkpoint inhibitor [ICI] therapy of nivolumab in this example) produces varied response in patients, which could be partially explained by the tumor expression of programmed-death-ligand 1 (PD-L1).9 Therefore, PD-L1 expression evaluation by a pathologist has become the standard procedure in patient selection for treatments and PD-L1 has been approved as a companion diagnostic biomarker for several drugs.
However, heterogeneity in PD-L1 expression (in lung cancer and over time) and limited biopsy samples could yield biased information for clinical evaluation and could lead to ineffective treatment. Some patients with high PD-L1 expression did not show positive response to treatment, while some with low PD-L1 expression did. Current discovery of 3D imaging technology (NSCLC 3D PD-L1 profiling) increases the sampling rate with continuous scanning of thick tissue, resulting in 100 times more pathology data from each patient, and tracing heterogeneous PD-L1 expression to give new parameters to support evaluation by the pathologists. After 3D image information is retrieved, the tissue may be reused for 2D hematoxylin and eosin (H&E) and immunohistochemistry (IHC) examination, as well as genetic testing by polymerase chain reaction (PCR). In other words, it allows the same clinical sample to be subsequently examined by multiple diagnosis methodologies and provides more integrated parameters to physicians with minimal changes in the clinical workflow.
Additional Clinical Implications Of This Type Of New Solution
Precision diagnosis could not only help cancer patients be matched with the most suitable treatment but also increase the success rate of new therapeutics. Continuing with the lung cancer example, IVD grade PD-L1 IHC assay is the only biomarker approved for patient screening in companion with selective indications of several ICI drugs such as pembrolizumab (Merck), nivolumab (BMS), and atezolizumab (Roche). General cutoff values of tumor proportion scores (TPS) are set as 50%, 10%, and 1% for treatment guidance.
Since PD-L1 expression in tumor tissue is heterogeneous, a pathologist’s choice of a single slide anywhere in the tissue for conventional IHC assay would often give a different measurement, consequently influencing the treatment plan. Some doctors think that tumor infiltrating T cells or other biomarkers in the tumor microenvironment should be taken into consideration for comprehensive diagnosis. However, it is even more challenging for pathologists to capture enough T cells or other biomarkers co-located in the same plane in the thin single slide (~4 μm in thickness), not to mention a complicated calculation.
PD-L1 positive rate for diagnosis is around 20-30%; those patients diagnosed with high PD-L1 will go for ICI treatment with insurance coverage (in Taiwan where the study took place).9 But those primarily diagnosed with a low PD-L1 would need a second chance for thorough inspection of the tissue for the PD-L1 signature to see if they are truly PD-L1 negative. Using computer-assisted PD-L1 quantitation method (i.e., an AI model for tumor cell recognition and PD-L1 positive cell identification) co-developed with pathologists to support diagnosis, a 3D PD-L1 TPS score was generated to categorize results with maximum, minimum, and average of values and compared to the 2D TPS scores and classification. Reclassifications lead to two cases passing the threshold for receiving treatment with medication concordance. This study supports the potential value of using an extra dimension in improving precision in PD-L1 measurement and applying 3D pathology evaluation to the FFPE specimen, which helps address the problem of uncertainties in diagnosis due to insufficient sampling (hence, inconsistencies) and allows patients to have the choice to receive treatment should the improved test results indicate it is appropriate. With further combination with other pathological data and leveraging AI/ML technologies, the 3D pathology based panoramic platform might aid in personalized or more tailored medicine in the future.
n addition to having approved many AI/ML-based medical technologies via 510(k) clearance, premarket approval, and de novo pathways,4 the U.S. FDA, Health Canada, and the U.K.’s MHRA have joined efforts to identify 10 guiding principles that will inform the development of good machine learning practice (GMLP) to promote safe, effective, and high-quality devices that use AI/ML.10 This is to address the unique nature of these products (i.e., complex, iterative, and data-driven) and the challenges/special considerations associated with them. Accuracy and consistency of diagnosis, prevention of bias to improve healthcare outcome and reduce costs, as well as articulating the safety and effectiveness of the products are key considerations for companies developing these types of solutions.11
In Part 2 of this article series, we will discuss considerations for collaboration between an AI/ML tech company and a hospital to develop and validate a technology solution for improving healthcare outcomes. Some potential barriers to and options for implementation and adoption of this type of solution in healthcare settings will also be discussed.
- Parwani, A.V. (2021). Next Generation Diagnostic Pathology: Use of Digital Pathology and Artificial Intelligence Tools to Augment a Pathological Diagnosis. Diagnostic Pathology, 14:138. Retrieved from https://doi.org/10.1186/s13000-019-0921-2
- Yang, D. Y., Fineberg, H. V., & Cosby, K. (2021). Diagnostic excellence. JAMA, 326(19):1905-1906.
- Adler-Milstein, J., Chen, J. H., & Dhaliwal, G. (2021). Next Generation Artificial Intelligence for Diagnosis: From Predicting Diagnostic Labels to “Wayfinding.” JAMA, 326(24):2467-2468.
- Benjamens, S., Dhunnoo, P., & Meskó, B. (2020). The State of Artificial Intelligence-Based FDA-Approved Medical Devices and Algorithms: An Online Database. npj Digital Medicine 3:118. Retrieved from https://doi.org/10.1038/s41746-020-00324-0
- Hey, S. P., & Kesselheim, A.S. (2016). Countering Imprecision in Precision Medicine: Better Coordination is Needed to Study Complex Interventions. Biomedical Research 353(6298):448-449.
- Schiff, G.D., Martin, S. A., Eidelman, D., Volk, L., Ruan, E., Cassel, C., Galanter, W., Johnson, M., Jutel, A., Kroenke, K., Lambert, B., Lexchin, J., Myers, S., Miller, A., Mushlin, S., Sanders, L., Sheikh, A. (2018). Ten Principles for More Conservative, Care-full Diagnosis. Annals of Internal Medicine 169(9):643-645. Retrieved from https://doi.org/10.7326/M18-1468.
- Ahsan, M. M., Nazim, R., Siddique, Z., & Huebner, P. (2021). Detection of COVID-19 Patients from CT Scan and Chest X-ray Data Using Modified MobileNetV2 and LIME. Healthcare 9:1099. Retrieved from https://doi.org/10.3390/healthcare9091099
- Matheny, M. E., Whicher, D., Israni, S. T. (2019). Artificial Intelligence in Health Care: A Report From the National Academy of Medicine. JAMA 323(6):509-510. Retrieved from https://doi:10.1001/jama.2019.21579
- Lin, Y.-Y., Wang, L.-C., Hsieh, Y.-H., Hung, Y.-L., Chen, Y.-A., Lin, Y.-C., Lin, Y.-Y., & Chou, T.,-Y. (2022). Computer-Assisted Three-Dimensional Quantitation of Programmed Death-Ligand 1 in Non-Small Cell Lung Cancer using Tissue Clearing Technology. Journal of Translational Medicine 20:131. Retrieved from https://doi.org/10.1186/s12967-022-03335-5
- US Food and Drug Administration (2021). Good Machine Learning Practice for Medical Device Development: Guiding Principles. Retrieved from https://www.fda.gov/media/153486/download
- Smith, A. (2021). 4 Lessons For AI In Medtech: Case Studies From Breast Cancer Detection. Retrieved from Med Device Online: https://www.meddeviceonline.com/doc/lessons-for-ai-in-medtech-case-studies-from-breast-cancer-detection-0001
About The Authors:
Peiyi Ko, Ph.D., CHFP, is founder and consultant at KoCreation Design LLC. Since 2017, she has researched and promoted integrated quality and adoption of technology by the life science industry. She has executed and managed projects, presented at conferences and university classes, and led workshops. She obtained her Ph.D. from the University of California, Berkeley, where she also completed the Engineering, Business, and Sustainability Certificate and the Management of Technology Certificate programs in 2011. You can reach her at firstname.lastname@example.org and connect on LinkedIn.
Margaret Dah-Tsyr Chang, Ph.D., is founder and Chief Strategy Officer of JelloX Biotech Inc. She has expertise in molecular biology and protein engineering supporting biotech-academic collaboration and biomedical translation from university to industry. She has executed and managed research and education projects, offered courses, promoted IP engineering and technology licensing, and organized conferences at National Tsing Hua University, Taiwan. She received her Ph.D. from the Department of Chemistry, Johns Hopkins University, and immediately established the first undergraduate molecular biology laboratory in Taiwan in 1993. You can reach her at email@example.com.