Now artificial intelligence discovers tumors and recommends treatments

A new artificial intelligence model could revolutionize the world of oncology. It is called Chief (an acronym for Clinical Histopathology Imaging Evaluation Foundation), and it is the first multipurpose device dedicated to medicine, capable not …

Now artificial intelligence discovers tumors and recommends treatments

A new artificial intelligence model could revolutionize the world of oncology. It is called Chief (an acronym for Clinical Histopathology Imaging Evaluation Foundation), and it is the first multipurpose device dedicated to medicine, capable not only of making diagnoses, but also of directing oncologists to the choice of therapies most suited to the characteristics of each individual tumor, predicting the chances of success of treatments and how long patients will survive. All of this – as described in Nature by its creators – with an effectiveness far superior to that of similar programs developed in the past, and almost comparable to that of a flesh-and-blood specialist.

The ChatGpt of oncology

Chief was created by a team of biomedical informatics experts from Harvard, and is designed to overcome the one-dimensional approach with which AI diagnostics have been developed to date. As a rule, these programs have an extremely limited scope of use: a single type of tumor, a single function (usually diagnosis starting from biopsy images) and the ability to only analyze images processed in a very specific way.

Chief is the exact opposite: he can do a bit of everything, and better than his predecessors. “Our ambition was to create an agile and versatile platform like ChatGpt, capable of carrying out a multitude of tasks related to tumor assessment,” explains Kun-Hsing Yu, researcher leading the team that developed Chief. “Our model has proven to be extremely useful in many different areas, relating to diagnosis, evaluation of prognosis and response to therapies, and for multiple types of cancer.”

An effectiveness never seen before

To give it its incredible diagnostic capabilities, the researchers fed Chief over 15 million high-resolution histopathological images, which the researchers divided into areas of interest, and another 60 thousand whole images. One strategy that has helped the AI ​​is to develop a holistic interpretation of the photographs submitted to it. The training was also carried out on images of tissues from 19 different anatomical locations, so that Chief was able to work on many types of tumors, on samples taken during tumor resection and on biopsies.

Once you have the desired multipurpose program, the time has come to put it to the test. To do this, Harvard researchers asked Chief to evaluate 19,400 histopathological images from 32 databases and collected in 24 different hospitals around the world. And in fact, the average success rate in identifying cancerous lesions within the samples was 94 percent, with peaks of 96 percent reached in the case of biopsies for some types of neoplasms.

Not just diagnosis

Identifying tumors isn’t Chief’s only ability. The program has in fact proven capable of predicting the molecular profile of tumors starting from their appearance, and thus recommending the best drug with which to treat them among those available. It has demonstrated the ability to predict patients’ short- and long-term survival chances. And it even identified some previously unknown histopathological features with prognostic value.

“If it is further validated, our approach could also help to early identify patients who can benefit from experimental treatments that target specific molecular alterations – explains Yu – a possibility that is not uniformly available today throughout the world”.

Compared to other artificial intelligences dedicated to oncology, Chief’s results were on average 36 percent better. But its developers hope to do even more in the future: they are currently working to make Chief capable of analyzing images of rare tumors and non-oncological diseases, to train him to work on precancerous lesions, improve his ability to evaluate the aggressiveness of tumors, and calculate the potential benefits of new experimental treatments, comparing them with those of standard therapies.