These days, it seems the concept of artificial intelligence is unavoidable. What once was a dystopian idea, a glimpse into an automated future far, far away, has seemingly arrived on our doorstep. In its current form, AI programs like ChatGPT and DALL-E are far from sentient beings, but the rate of progress these sophisticated algorithms have developed is alarming for some. But what positives can this technology bring to the table? We caught up with SCI member Ruijiang Li, PhD, associate professor of radiation oncology in the division of medical physics, to gain insight. He and his colleagues in the Integrative Imaging & Molecular Diagnostics Lab have been developing a powerful approach to medical image analysis, utilizing a deep-learning program to predict treatment response and patient outcomes based on their routine CT and MRI scans, as well as digital pathology slides.
“Unlike the traditional approach, [the deep-learning model] does not rely on human input or expert knowledge. Instead, it learns informative patterns from the data that are relevant to the problem at hand,” Li explained. “It’s fully automatic. We supply the data, and it provides the output; the desired response.”
By incorporating this revolutionary approach to cancer imaging into the clinical workflow, Li anticipates that what once took hours of human labor can be done in the click of a button, so to speak. From there, he hopes to successfully “teach” the algorithm to take data from routine cancer care and use it to create a personalized treatment approach.
“There has been much progress in using deep-learning to improve cancer screening and detection, thanks to readily available large datasets. In my lab, we’re more interested in the downstream task along the cancer care pathway, which is prognosis and treatment response prediction,” he said.
The drawback with this approach is the amount of data required to train the model what to identify. Because of the model’s size, it has millions of parameters in a typical network, requiring a very large data set to train a reliable model.
A solution to this problem is “federated learning,” a decentralized approach to training machine learning models by leveraging datasets from multiple institutions of the same caliber, without compromising patient privacy.
Li and his team are also exploring a concept known as “biology-guided deep-learning,” where the deep-learning model is trained to not only predict treatment outcomes for the patient, but is able to identify the underlying tumor biology itself. To do so, the system was fed information based on patient medical history and prior knowledge of cancer biology; established biomarkers, such as the status of the immune and stromal microenvironment, to name a couple. Classifying tumor subtypes is imperative to cancer treatment, as some subtypes of tumors will respond differently to different types of treatment. By being able to correctly predict the response to different courses of treatment, such as immunotherapy, the model was able to identify the corresponding patient subset group that the treatment would be most beneficial to. Using AI to streamline the process of finding that patient subset group is a massive innovation that will enable personalized cancer care.
“This is called multi-task learning. Because of that design, the model generalizes better than traditional deep-learning, and it’s more biologically interpretable,” Li remarked.
In its current state, however, AI is simply a tool to help aid human clinicians, who will have ultimate discretion for choosing treatment courses.
“AI is an adjunct tool, not the primary diagnostic tool,” said Dr Li. “A main advantage is that it can reduce diagnostic errors and improve cancer detection rate.” He maintains that human expertise is the key to efficiently functioning AI.
Looking forward, Li predicts that AI will play a large role in the field of oncology, with regards to both predicting treatment outcome and cancer screening and detection, but that’s still years away. This new technology needs to undergo rigorous testing and real-world evaluation before it can reliably weigh in and make decisions on a patient’s health.
“The biggest issue is still validation. These models work very well on retrospective data, but ultimately, they need to be tested in prospective studies and ideally show improved outcomes in randomized trials. Things are moving along, but it will take some time before we get there,” Li said.
Visit the Li Lab to learn more about Li’s current research.
