Scientists at Harvard Medical School have designed an artificial intelligence (AI) model with capabilities similar to ChatGPT, capable of performing a wide range of diagnostic tasks in different types of cancer.
The system, called CHIEF (Clinical Histopathology Imaging Evaluation Foundation)was designed to overcome the limitations of current models, which tend to focus on specific tasks or narrow cancer types, according to an article published last September in the magazine Nature.
CHIEF has demonstrated effectiveness in 19 different types of cancermarking a significant advance in the use of artificial intelligence for medical diagnosis.
“Our ambition was to create an agile and versatile AI platform that could perform a wide range of cancer assessments,” he said. Kun-Hsing Yusenior author of the study and assistant professor of biomedical informatics at the Blavatnik Institute at Harvard Medical School.
“Our model turned out to be very useful in tasks related to detection, prognosis and response to treatment in multiple types of cancer”he added.
CHIEF analyzes digital images of tumor tissues, identifying cancer cells and predicting molecular profiles with greater accuracy than current systems. Besides, predicts patient survival in different types of cancer and evaluates characteristics of the tumor microenvironment, elements of the surrounding tissue that influence the response to treatments such as surgery, chemotherapy, radiotherapy and immunotherapy.
The team highlighted that this new AI tool also identified previously unknown tumor characteristics that are related to patient survival, which could open new lines of research and treatment.
“If validated and deployed widely, our approach could early identify “patients who could benefit from experimental treatments targeting specific molecular variations”Yu explained, underscoring the importance of these capabilities in regions with limited access to advanced treatments.
The model was initially trained with 15 million unlabeled images and later with 60,000 complete images of tissues from 19 types of cancerincluding lung, breast, prostate, stomach, kidney, brain, skin and pancreas. This approach allowed CHIEF to analyze images holistically, relating changes in specific areas to the overall context of the tissue.
During testing, the tool was evaluated with more than 19,400 images from 32 independent data sets and 24 hospitals from different regions of the world. The model outperformed other advanced systems by up to 36% in key tasks such as the detection of cancer cells, the identification of the origin of the tumor, the prediction of clinical outcomes and the detection of genes associated with response to treatment.
One of the notable features of CHIEF is its ability to maintain accuracy regardless of how the tumor cells are obtained, whether by biopsy or surgery, and the method used to digitize the samples. This makes it a tool adaptable to various clinical settings, a significant advance over current models, which tend to be less effective outside of specific conditions.
The development of CHIEF builds on previous research by Yu’s team, which demonstrated the feasibility of using AI in the analysis of colon and brain tumors. According to the researchers, this new model represents a step towards the more widespread use of AI to improve the accuracy and efficiency in cancer diagnosis and treatment.
In five biopsy data sets collected from independent cohorts, The AI system achieved 96 percent accuracy in multiple cancer types, including the esophagus, stomach, colon and prostate. When the researchers tested it on previously unseen slides of tumors surgically removed from the colon, lung, breast, endometrium and cervix, the model performed with more than 90 percent accuracy.
For its part, in all types of cancer and in all groups of patients under study, CHIEF distinguished patients with longer-term survival from those with shorter-term survival. According to Harvard, CHIEF outperformed other models by 8 percent; and in patients with more advanced cancers, it outperformed other models by 10 percent. In total, its ability to predict high versus low risk of death was tested and confirmed in patient samples from 17 different institutions.
