New Harvard AI Model Detects Cancer With 96% Accuracy
A novel AI model known as CHIEF (Clinical Histopathology Imaging Evaluation Foundation), developed by researchers at Harvard Medical School, has been very accurate in diagnosing and forecasting the course of certain cancer types.
The study claims that CHIEF works better than current AI systems, reaching up to “96% accuracy” in the diagnosis of 19 distinct cancer kinds. The researchers compare CHIEF’s adaptability to that of ChatGPT, a language model that has drawn a lot of interest because of its capacity to handle a variety of jobs.
Instead of taking the generalist approach found in more conventional models like GPT-4V or LlaVA, CHIEF is essentially a highly specialized AI vision model—one that can comprehend visual inputs—trained to be extremely detailed in images of cancer cells.
Therefore, CHIEF was trained on a large multimodal dataset that included 15 million unlabeled photos and 60,000 whole-slide images of tissues from 19 different anatomical sites, rather than being trained to recognize generic elements like “cats” or “oranges.”
According to the study, “CHIEF extracted microscopic representations useful for cancer cell detection, tumor origin identification, molecular profile characterization, and prognostic prediction through pretraining on 44 terabytes of high-resolution pathology imaging datasets.”
The strategy appears to be more effective than expected. According to research senior author Kun-Hsing Yu, “our goal was to develop a quick, flexible AI platform that can carry out a variety of cancer evaluation tasks, similar to ChatGPT.” “Our model proved to be very helpful in a variety of tasks pertaining to the detection, prognosis, and response to treatment of various cancers.”
CHIEF surpassed state-of-the-art AI techniques by up to 36.1% on all tasks when tested on over 19,400 photos from 32 separate datasets gathered worldwide. It also was better at distinguishing between patients with high and low survival rates and could accurately offer information about various tissue samples that were tested.
To increase CHIEF’s accuracy, the researchers intend to train it on pictures of non-cancerous illnesses, rare disorders, and pre-malignant tissues. To improve the model’s ability to detect cancer aggressiveness and forecast the results of innovative treatments, they also expect feeding it additional data.
An Expanding Role in Cancer Detection and Beyond. For some time now, scientists have been using AI to improve cancer and other disease detection, diagnosis, and therapy.
For instance, Cambridge researchers unveiled EMethylNET, an AI model that has a 98% success rate in identifying 13 different forms of cancer using DNA information from tissue samples. Trained on over 6,000 tissue samples, EMethylNET demonstrates how AI may detect cancer early by detecting DNA methylation, which is a key factor in the development of cancer.
An earlier model called CancerGPT (I’m not making that word up) predicted how drug combinations could affect cancer patients’ unusual tissues using a huge language model. It proved that when structured data and samples are limited, pre-trained models can be quite helpful. CancerGPT can provide important insights by generalizing predictions and using previous medical research, but researchers were still worried about possible AI hallucinations.
Google and iCAD collaborated to improve cancer screening with AI. Their AI-powered approach improved accessibility to life-saving breast cancer screenings and exceeded professional radiologists in accuracy, providing a workable solution in the face of a global radiology deficit.
Lastly, brain surgeons are using another AI tool called Sturgeon to help them diagnose malignancies in the central nervous system in real time with 90% accuracy.
CHIEF is open-source and may be downloaded from the project’s Github page, allowing researchers or anybody else to use it locally and add their own photos.