A New Artificial Intelligence Tool for Cancer
Researchers at Harvard Medical School have developed a groundbreaking AI model that demonstrates remarkable versatility in diagnosing various cancers. The new AI system, introduced in a September 4 article in Nature, represents a significant advancement over existing cancer diagnostic technologies.
Unlike current AI models that are typically designed for narrow tasks—such as detecting the presence of cancer or predicting genetic profiles of tumors—the new model is capable of performing a broad range of diagnostic functions. It was evaluated across 19 different types of cancer, showing a level of flexibility akin to that of large language models like ChatGPT.
The model, known as CHIEF (Clinical Histopathology Imaging Evaluation Foundation), is believed to be the first of its kind to predict patient outcomes and validate these predictions across diverse international patient populations. This sets it apart from other AI models which have primarily focused on pathology images without broader validation.
"Our goal was to create a nimble and versatile AI platform that could handle various cancer evaluation tasks," said Kun-Hsing Yu, the study's senior author and an assistant professor of biomedical informatics at Harvard Medical School's Blavatnik Institute. "Our model has proven to be extremely effective in cancer detection, prognosis, and treatment response across multiple cancer types."
CHIEF operates by analyzing digital slides of tumor tissues, where it identifies cancer cells and predicts molecular profiles based on the features visible in these images. It excels in forecasting patient survival and pinpointing tumor microenvironment features that influence treatment responses, such as those to surgery, chemotherapy, radiation, and immunotherapy. The tool has even revealed new insights by identifying tumor characteristics previously unknown to affect patient survival.
These findings highlight the potential of AI to significantly enhance cancer diagnosis and treatment, including identifying patients who may not respond to standard therapies. "If our approach is validated further and widely implemented, it could help identify cancer patients who might benefit from experimental treatments targeting specific molecular variations," Yu added.
The development of CHIEF builds on Yu's prior work with AI systems for colon and brain cancer. The new model was trained on 15 million unlabeled images, followed by detailed training on 60,000 whole-slide images covering 19 cancer types. This dual-layer training allowed CHIEF to interpret images in a comprehensive context rather than focusing on isolated regions.
The model was evaluated using over 19,400 whole-slide images from 32 datasets across 24 hospitals worldwide. CHIEF outperformed current state-of-the-art AI methods by up to 36 percent in tasks such as cancer cell detection, tumor origin identification, and predicting patient outcomes. Its performance remained consistent across various methods of obtaining and digitizing tissue samples, demonstrating its versatility in different clinical settings.
Key achievements of CHIEF include:
Nearly 94 percent accuracy in cancer detection across 15 datasets with 11 cancer types.
Successful prediction of patient survival based on initial diagnostic images.
Discovery of new patterns related to tumor aggressiveness and patient survival, visualized through heat maps generated by the AI.
Looking ahead, the research team plans to enhance CHIEF's capabilities by:
Training the model with images of rare diseases and non-cancerous conditions.
Including pre-malignant tissue samples.
Incorporating additional molecular data to improve detection of cancer aggressiveness.
Expanding the model's ability to predict responses to novel cancer treatments.
These steps aim to further refine CHIEF's diagnostic power and adaptability, making it a valuable tool in the ongoing effort to advance cancer treatment and care.
Source: https://hms.harvard.edu/news/new-artificial-intelligence-tool-cancer