In a recent study, researchers developed and evaluated an AI-based pathology model called Prov-GigaPath. According to the researchers, this is the first whole-slide pathology foundation model for diagnosing cancer cells trained on large data sets from real-world cases.
Computational pathology aids in transforming cancer diagnosis, helping professionals identify disease subtypes, stages, and possible progress. In many studies, machine learning and deep learning have shown better results for early cancer detection of various types.
Also read: Scientists Develop AI-Powered Model to Enhance Cancer Immunotherapies
Providence Health Systems and the University of Washington conducted the latest research study, published in the Journal Nature. A number of Microsoft’s in-house teams also collaborated to facilitate the research.
Prov-GigaPath Diagnoses Cancer
Prov-GigaPath builds on the whole-slide imaging method, which is widely applied to cancer evaluation and diagnosis.
In the whole-slide imaging technique, a microscopy slide of a tumor image is turned into a high-resolution digital image. These whole-slide images contain critical information that helps understand the tumor microenvironment.
“Prov-Path is more than five times larger than TCGA in terms of the number of image tiles and more than two times larger than TCGA in terms of the number of patients.” Nature.
Prov-GigaPath is trained on a large dataset called Prov-path from Providence Health Network, which has 28 cancer centers. The dataset has over 1.3 billion image tiles from 171,189 actual microscopy slides. The slides were developed during biopsies and resections of more than 30,000 patients and cover 31 major tissue types.
The Prov-Path dataset also contains data on cancer staging, related pathology reports, genome mutation profiles, and histopathology findings. Together, these diverse data parts provide a better understanding of the conditions for the model.
GigaPath Enhances Gigapixel Slides Identification
GigaPath is a new vision transformer that Prov-GigaPath uses to evaluate gigapixel pathology slides. A complete slide becomes a series of tokens when the image tiles are used as visual tokens. In order to simplify complicated patterns for sequence modeling, the vision transformer is a neural architecture.
The point is that a conventional vision transformer cannot be applied directly to digital pathology because of the sheer number of tiles in each microscope slide. In the case of Providence data, the number of slides can be as high as 70,121. The researchers noted that,
“To address this problem, we leverage dilated self-attention by adapting our recently developed LongNet method.”
Many function-altering gene mutations are involved in cancer progression, which can be screened for both cancer diagnosis and prognosis. The study noted that despite the significant decrease in the cost of sequencing, there are still healthcare gaps. Access to tumor sequencing worldwide is said to be the primary factor for the said gap.
The researchers highlighted that predicting tumor mutations from pathology images can help select treatment methods and personalized medication.
Researchers Compare Pathology Models
Digital pathology has computational challenges, as standard gigapixel slides are usually thousands of times larger than traditional natural images. Conventional vision transformers have limitations and struggle to handle such gigantic images because computational requirements increase with such amounts of data.
Also read: AI Tool Predicts Immune Responses in Fight Against Cancer
Another point is that previous research in digital pathology did not leverage the interdependencies across different image tiles in each microscopy slide. This ignorance of connecting the interdependencies led to eliminating slide-level context, which is crucial for many applications, such as tumor microenvironment modeling.
Researchers compared Prov-GigaPath against other publicly available pathology foundation models such as HIPT, Ctranspath, and REMEDIS for the study. Researchers found that Prov-gigaPath showed better performance on 25 out of 26 tasks, as the study noted that,
“Prov-GigaPath attained an improvement of 23.5% in AUROC (a performance measure for classification models) and 66.4% in AUPRC (a measure useful when dealing with imbalanced datasets) compared with the second-best model, REMEDIS.”
Cancer can be a life-threatening disease, and it costs millions of lives every year. As Thomas Fuchs, the co-founder and chief scientist at digital pathology provider Paige, told CNBC in an interview, “You don’t have cancer until the pathologist says so. That’s the critical step in the whole medical edifice.”
As we know, conventional pathology techniques have assisted in the diagnosis of diseases because they largely rely on looking at tissue samples under a microscope. However, with technology and artificial intelligence at hand, practices are changing, and the process of identifying and classifying cancers has accelerated. Most AI pathology models leverage the same technique of examining microscopy slides but in a digital way.
Cryptopolitan reporting by Aamir Sheikh
