CHIEF Model Transforms Cancer Evaluation with Enhanced Generalizability Across Diverse Histopathology Samples

By Dr. Chinta SidharthanReviewed by Lily Ramsey, LLMSep 16 2024

The diagnosis of cancers and the classification of tumor subtypes rely on histopathological imaging and analysis. Artificial intelligence (AI) models are being increasingly used to optimize the diagnosis of cancers using histopathological imaging.

In a recent study published in Nature, a team of scientists addressed the limited generalizability of some of the AI models in analyzing images from different populations and digitization methods by designing a general-purpose machine learning framework called Clinical Histopathology Imaging Evaluation Foundation or CHIEF that can extract various features from pathology images for cancer diagnosis and evaluation.

​​​​​​​Study: A pathology foundation model for cancer diagnosis and prognosis prediction. Image Credit: Komsan Loonprom/Shutterstock.com

Background

The use of AI tools in histopathological analyses for cancer diagnosis is advancing rapidly. Traditional AI models are trained for specific tasks, such as identifying cancer cells or predicting treatment outcomes.  

However, these models require large training datasets, and their results are often inconsistent across different imaging methods or tissue types. They are generally adapted from computer vision models originally designed to identify macroscopic objects.

Self-supervised learning is a more flexible method that uses unlabeled data to train AI models and has been observed to perform better across different tasks.

However, despite recent advances in self-supervised learning models, their focus on narrow tasks and lack of generalizability continue to present challenges in the wide use of AI models for cancer diagnosis.

About the Study

In the present study, the researchers used a large dataset consisting of over 60,000 whole slide images from 14 different study groups to train and develop the CHIEF model. These whole slide images spanned 19 different types of cancers from a wide variety of anatomical regions such as the brain, breast, lung, liver, etc.

The training of the CHIEF model was conducted in two stages. The first stage involved the model undergoing self-supervised learning using unlabeled data to learn how to extract patch-level features from all the slide images.

Patch-level feature extraction involves the breaking down of an image into smaller, fixed-size segments or "patches." Instead of analyzing the entire image at once, the model focuses on these patches to extract localized features.

The second stage of the training involved integrating these patches using attention modules and weakly supervised learning to create a global representation of the pathology images.

The second step allowed CHIEF to understand the larger context of the image without depending on detailed annotations, which are not available in every scenario.

The model also incorporated test data along with image data to improve the ability of the model to analyze histopathological images.

The neural network model used in the image encoder in CHIEF was CTransPath, which has previously been trained using 15 million image patches. A preexisting model known as CLIP, which has been trained on image datasets and their associated text datasets, was used in the text encoder.

Techniques to clean images and remove irrelevant background data were also incorporated into the model.

The performance of the model was then validated using a variety of tasks, such as cancer cell detection, identification of tumor origin, and prognosis or survival prediction. Datasets from 24 hospitals were used to validate the performance of the model in these tasks.

Major Findings

The study demonstrated that CHIEF outperformed the other models, such as ABMIL, CLAM, and DSMIL in cancer detection across 15 independent datasets spanning 11 cancer types.

Assessment of the model’s area under the receiver operating characteristic curve (AUROC), which is an indicator of overall performance, reported that the performance of CHIEF was 10% to 14% better than that of other models, with an AUROC of 0.9397.

Furthermore, the detection of cancerous regions by CHIEF at the pixel level aligned closely with the annotations provided by the pathologist. CHIEF was also able to predict tumor origins with high accuracy, identifying various cancer-associated genetic mutations such as those in tumor protein P53 gene in the case of low-grade gliomas, and BAP1 or breast cancer gene 1 (BRCA1)-associated protein one mutations in uveal melanomas.

The ability of CHIEF to predict mutation related to cancer therapies approved by the United States (U.S.) Food and Drug Administration (FDA) also highlighted the model’s potential use in personalized treatment planning.

The researchers also found that CHIEF outperformed other models in the area of survival prediction by using histopathological images to distinguish between patients with longer and shorter survival outcomes accurately.

Conclusions

Overall, the findings suggested that CHIEF was a versatile tool for accurately detecting cancers across various tissue and cancer types and for predicting genomic profiles and survival outcomes.

The weakly supervised self-learning framework provides the model with the advantage of generalizability across a diverse range of cancers and tissue types.