A new artificial intelligence model designed by researchers at Harvard Medical School and National Cheng Kung University in Taiwan could provide important help to doctors and patients with colon cancer – the world’s second deadliest cancer – who have to decide for treatments, but also to know the survival rates according to the special characteristics of their disease. The new discovery is published in Nature Communications.
By looking at images of cancer cells, the new tool is able to accurately predict how aggressive a tumor is, how likely the patient is to survive with or without a recurrence of the disease, and what the optimal treatment might be for them.
A tool that can provide answers to such key questions could help doctors and patients manage serious disease more effectively, saving millions of lives.
The tool, called MOMA, was trained on data from 2,000 colon cancer patients from several national studies. During training, the researchers fed the model information about the patients’ age, sex, and cancer stage. They also gave him information about the genomic, epigenetic, protein and metabolic profiles of the tumors.
The researchers then showed the model tumor images and asked it to look for visual markers related to tumor types, genetic mutations, epigenetic alterations, disease progression and patient survival.
The researchers then tested how the model would perform in the “real world” by giving it a set of tumor images of various patients it had never seen before. They compared its performance with actual patient outcomes and other available clinical information.
The model accurately predicted patients’ overall survival after diagnosis. It also accurately predicted how an individual patient might respond to different treatments. In both of these areas, the tool outperformed pathologists as well as available artificial intelligence models.
The researchers said the model will undergo periodic upgrades as the science evolves and new data emerges.
The study team cautions that the tool is intended to augment, not replace, human experience.
“Our tool performs tasks that pathologists cannot do,” said study co-leader Kun-Hsing Yu, assistant professor of biomedical informatics at the Blavatnik Institute at HMS. “What we expect is not a replacement for the expertise of human medicine, but that this approach will enhance current clinical practice of cancer management.”
The researchers caution that each patient’s prognosis for survival depends on many factors, and no single tool provides accurate predictions. They add, however, that the new model could help clinicians more accurately choose the appropriate treatment for each patient case.
The researchers point out that the new tool goes beyond many AI models, which mainly perform tasks that replicate or optimize human expertise. Instead, the new tool detects and interprets visual patterns invisible to the human eye.
The new model offers unprecedented levels of detail, which are not visible to the eye. For this reason, it is capable of providing timely support in resource-constrained environments, the researchers said.
They added, however, that before it can be used by clinics and hospitals, it should be tested in a randomized trial that evaluates the performance of the tool in real patients, over time, after the initial diagnosis.
The model accurately identified image features associated with differences in survival.
For example, he identified three characteristics that portend a worse prognosis
- Greater cell density in a tumor
The presence of connective supporting tissue around the cancer cells, known as the layer.
- Interactions of cancer cells with smooth muscle cells
The model also identified patterns that showed which patients were more likely to live longer without cancer recurrence.
- The tool also accurately predicted which patients would benefit from a class of treatments known as immune suppressants.
These are treatments that work for many patients, but there are also many who show no measurable benefit and experience serious side effects.