Is AI the Missing Piece Precision Medicine Has Been Waiting For?

Artificial intelligence (AI) is now relevant across nearly every area of research, biomedical and otherwise. And while pathology is no exception, it may be one of the last frontiers—it’s estimated that only five to ten percent of hospitals have transitioned from glass slides to digital images. Here, we’ll explore how close we are to realizing AI-guided personalized care, along with the challenges and benefits associated with integrating this new technology into medicine.

Pathology is central to nearly all clinical advancements

Pathology is critical to the development of new drugs. Historically, the field has contributed to preclinical research through target identification and delineation of drug mechanisms of action, while also assessing pharmacodynamics and toxicology. Pathology also connects drug discovery to translational efforts and clinical efficacy, offering insight into disease pathophysiology, patient treatment and response, and more. To illustrate, pathology-based observations are increasingly used to determine drug efficacy across diseases. A drug development trial for non-alcoholic steatohepatitis used a histological endpoint as the basis for moving to the next phase. In another study, the use of pathological complete response (pCR) was used as an endpoint, and associated with improved long-term efficacy of combination chemotherapy and trastuzumab in patients with human epidermal growth factor receptor 2 (HER2)-positive breast cancer^.1,2

Setting the stage for AI

The increased emphasis on precision medicine and advances in technology have spurred the development of digital pathology-based approaches in recent years. In the past, pathologists focused on regions of interest (ROI) from tissue samples to make a diagnosis or treatment recommendation. Now, however, there are whole slide imaging (WSI) scanners, which digitally stitch together multiple images from a sample to generate the entire tissue section. These include the Philips IntelliSite Pathology Solution and Leica Aperio AT2 DX System, which are FDA-approved for the review of digital surgical pathology slides from biopsied tissue. These digitized tissue images enable international remote communication and review from multiple pathologists, leading to less inter-pathologist variability and allowing for the generation of an accessible, curated database with unlimited applications from the educational to the clinical.^3,2

The combination of WSI and newer multiplexed IHC, which can label several molecular markers simultaneously, can be used to decipher the spatial relationships of phenotypically distinct cell populations beyond ROIs and encompassing the entire tissue area. The vast and rich availability of information that can now be derived means there is also a need for reproducing interpretation of these complicated data sets and has led to the use of AI in pathology.

AI in the immuno-oncology realm

The tumor microenvironment yields endless clues about cancer, like its molecular underpinnings, relationship with the immune system, and what could potentially destroy it—but only if the vast amounts of complicated data can be accounted for and interpreted. The fusion of WSI with AI algorithms is a powerful means to decode this information. Quantitative analysis can be used to create high content data through multiplexing, an immunohistochemical method that shows the expression and localization of numerous markers within the context of the complex tissue environment and at the single-cell level. These capabilities not only prevent human counting fatigue and the potential for error, but decrypt correlates relevant to patient treatment. In fact, adding spatial metrics, or additional surrounding context, to IHC findings also enhances biomarker identification approaches. A recent meta-analysis showed that adding this additional layer of spatial information to IHC and immunofluorescence led to significantly better prediction of response of cancers to immune checkpoint inhibitors (ICI), compared to IHC alone, or gene expression profiling.⁴

AI depends on algorithms to train computers to perform a specific task, either with or without human supervision. Unsupervised learning models have already been used in breast cancer samples to differentiate between low- and high-grade tumors, as well as to evaluate morphological features to provide a score associated with overall survival. Still these AI approaches rely on the quality and quantity of data used to train the algorithm.¹ This could be an issue as there are often differences in slide preparation, histological methods, and scoring, between scientists and labs. Another obstacle is the initial capital and time investment. Although digital pathology and AI mean faster, more efficient workflows, the financial cost of equipment and software, alongside the time-consuming nature of algorithm development, can't be ignored.

In immuno-oncology, AI-guidance is already helping to match patients to the most promising treatments. In melanoma patients, adoptive T cell therapy has a greater than 50% success rate, but only in patients who can generate tumor-infiltrating lymphocytes (TIL). Feng et al utilized 7-color multispectral IHC and unsupervised hierarchical clustering to thoroughly analyze the immune environment in melanoma tumors. The authors postulate that the immune profile could be used to select for patients with TIL capability, as well as predict who will respond to ICIs.⁵

Pancreatic adenocarcinoma is a cancer with a notoriously poor survival rate, typically thought to be non-responsive to immunotherapies. However, the automated classification of immune cells and epithelial cells alongside single-cell level marker analysis, showed that the spatial distribution of cytotoxic T cells with relation to cancer cells is a marker of increased overall patient survival. The authors propose that analyzing the TME for T cells may help to identify which patients would most benefit from immune therapies, and argue that this kind of information should be included in standard tumor pathological scoring.⁶

Other clues to navigating particular tumors have to do with tumor purity (TP) assessments. TP refers to the percent of cancer cells versus immune- and other types of cells present in the tumor tissue. Low TP may indicate the upregulation of immunosuppressive pathways, and this is associated with a poor prognosis, as is the case with gastric cancer. AI assessment of TP was found to be more accurate compared to visual interrogation by pathologists, further helping clinicians to guide treatment selection such as immunotherapy.⁷

Real-world integration

In the clinical market space, digital pathology is believed to be the next major digital transformation. In 2021, the FDA approved Paige’s Prostate Detect, which combines digital pathology with AI. Two years later, the company launched its suite of breast and colon cancer digital pathology products and is seeking FDA approval. More recently, Paige announced a collaboration with Microsoft to create the world’s largest image-based AI models for digital pathology and oncology, which will be configured with billions of parameters, and up to 4 million digitized microscopy slides from a multitude of cancers.³

The Mayo Clinic is also currently part way through a multi-phase digital pathology initiative to replace glass slides for clinical diagnosis. The hope is that this will lead to increased collaboration in assessing treatment plans, help to train new doctors, and allow for the creation of AI-based algorithms.

The promise of AI in medicine

The most appealing aspects of AI may be its capacity for identifying abnormalities that elude human detection, and its potential for discovering novel biomarkers, applicable not only to diagnostics, but also to the discernment of novel drug targets. AI can also aid in stratification of patients and choosing optimal treatment regimens, particularly in complex areas such as immuno-oncology, where therapies continue to diversify. In this dynamic landscape, the melding of artificial intelligence and pathology not only enhances our understanding of disease mechanisms but also boosts the efficiency and success rates of drug discovery, clinical trial design, and patient outcomes.

References

1. Baxi V, Edwards R, Montalto M, Saha S. Digital pathology and artificial intelligence in translational medicine and clinical practice. Mod Pathol. 2022 Jan;35(1):23-32. 

2. Jackisch C, Hegg R, Stroyakovskiy D, Ahn JS, Melichar B, Chen SC, Kim SB, Lichinitser M, Starosławska E, Kunz G, Falcon S, Chen ST, Crepelle-Fléchais A, Heinzmann D, Shing M, Pivot X. HannaH phase III randomised study: Association of total pathological complete response with event-free survival in HER2-positive early breast cancer treated with neoadjuvant-adjuvant trastuzumab after 2 years of treatment-free follow-up. Eur J Cancer. 2016 Jul;62:62-75. 

3. Shafi S, Parwani AV. Artificial intelligence in diagnostic pathology. Diagn Pathol. 2023 Oct 3;18(1):109.

4. Lu S, Stein JE, Rimm DL, Wang DW, Bell JM, Johnson DB, Sosman JA, Schalper KA, Anders RA, Wang H, Hoyt C, Pardoll DM, Danilova L, Taube JM. Comparison of Biomarker Modalities for Predicting Response to PD-1/PD-L1 Checkpoint Blockade: A Systematic Review and Meta-analysis. JAMA Oncol. 2019 Aug 1;5(8):1195-1204. 

5. Feng Z, Puri S, Moudgil T, Wood W, Hoyt CC, Wang C, Urba WJ, Curti BD, Bifulco CB, Fox BA. Multispectral imaging of formalin-fixed tissue predicts ability to generate tumor-infiltrating lymphocytes from melanoma. J Immunother Cancer. 2015 Oct 20;3:47. 

6. Carstens JL, Correa de Sampaio P, Yang D, Barua S, Wang H, Rao A, Allison JP, LeBleu VS, Kalluri R. Spatial computation of intratumoral T cells correlates with survival of patients with pancreatic cancer. Nat Commun. 2017 Apr 27;8:15095. 

7. Gong Z, Zhang J, Guo W. Tumor purity as a prognosis and immunotherapy relevant feature in gastric cancer. Cancer Med. 2020 Dec;9(23):9052-9063.