Machine Learning Predicts Immunotherapy Response in Metastatic Melanoma
Knowing which patients are likely to respond to therapy will lead to more personalized clinical decision-making
Recent advances in AI and machine learning hold promising potential across all aspects of health care, from diagnosis to treatment planning and response evaluation. Machine learning can predict response to immunotherapy in patients with metastatic melanoma, potentially speeding effective care for an often deadly
disease, according to research supported by an RSNA R&E Foundation grant. Melanoma is the most serious skin cancer due to its potential to spread to other parts of the body like the brain, liver and lungs. Five-year survival rates are excellent if the disease is diagnosed and treated at an early stage but fall significantly once the cancer has metastasized.
Immune checkpoint inhibitors (ICI), a type of immunotherapy, have shown great promise in treating metastatic melanoma. ICI works by blocking proteins on immune cells called check[1]points. Binding of checkpoint and partner proteins sends an ‘off’ signal to T-cells which can prevent the immune system from attacking cancer cells.
Thus, through inhibiting this binding, ICI allows the immune system to attack cancer cells. Research has shown that ICI can significantly increase survival, but not all patients benefit from it.
“It’s very hard to predict which patients are good candidates for immunotherapy,” said Dania Daye, MD, PhD, who received an R&E Foundation Research Seed Grant. “We wanted to see if there are any imaging markers that we can use to predict which patients will respond to immunotherapy and which will not.”
Machine Learning Model Can Offer Personalized Medicine
Dr. Daye, who is an interventional radiologist and assistant professor at Massachusetts General Hospital and Harvard Medical School, both in Boston, and colleagues used an existing database of images from melanoma patients that included 79 patients with 168 metastatic liver lesions. All patients had arterial phase CT images one month prior to initiation of ICI.
Response to ICI was assessed on follow-up CT at three months using Response Evaluation Criteria in Solid Tumors (RECIST) criteria.
The researchers developed a machine learning algorithm and trained it to find patterns in the CT images that are not quantitatively distinguishable to the human eye. Further analysis helped select the most important predictive features and variables.
They then evaluated the ability of the machine learning-based model to predict responses to immunotherapy and compared it with levels of lactate dehydrogenase (LDH) prior to treatment. LDH levels are a known biomarker to predict advanced melanoma’s response to ICI.
The machine learning model that images and pretreatment LDH levels resulted in better predictions of cancer response to ICI than models that only use radiomics features or LDH levels alone. Analysis revealed 15 features that are associated with a response to ICI.
“This is personalized medicine at its finest,” Dr. Daye said. “Being able to do pre-procedural or pre-treatment imaging and knowing which patients are more likely than others to respond to the therapy that we’re about to offer will lead to more personalized clinical decision-making.”
Dr. Daye and her colleagues are developing and training a new algorithm that uses deep learning to look at specific parts of the abdomen for patterns predictive of response to treatment. They also would like to conduct a prospective study to affirm the results from the retrospective work.
“This is personalized medicine at its finest. Being able to do pre-procedural or pre-treatment imaging and knowing which patients are more likely than others to respond to the therapy that we’re about to offer will lead to more personalized clinical decision-making.”
— DANIA DAYE, MD, PHD
“I’m very thankful for the R&E Foundation funding for this pilot project because it’s definitely allowed us to get the preliminary data that we need,” Dr. Daye said. “We’re currently looking for a larger grant to conduct a prospective validation on some of the findings from the study to see if we can get this to the clinic and to patients who actually need it,” she said.
Dr. Daye, who divides her time equally between clinical and academic work, credits the support of the R&E Foundation for propelling her career forward at critical junctures.
“I have been very lucky to be able to benefit from RSNA grants at every stage of my career,” she said. “I received a medical student grant, a resident grant and finally I had a pilot grant as a junior faculty. Each grant has allowed me to progress to the next stage of my career and provided me with great experience and preliminary data to move to the next project.”
Dr. Daye is passing those benefits on by mentoring a postdoctoral researcher in her lab who recently got a pilot grant and another student who is applying this year for a medical student grant. “Coming full circle to the other side and being able to mentor people is very, very rewarding,” she said.
For More Information
Learn about R&E Foundation funding opportunities.
Read previous RSNA News stories about machine learning in radiology: