“As pathologists, it’s our job to understand the story the tumor is telling us, but we have reached the limit of what our brains and eye can do.” Pathologists and computer scientists at USC are working together to better understand treatment-resistant cancer. Photo/iStock.
One of the biggest challenges facing cancer researchers today is predicting when cancer will resist chemotherapy. Hidden inside a cancer tumor’s dense network of genes, proteins, pathways, and molecules is a wealth of information. But often, tumors remain tight-lipped.
“As pathologists, it’s our job to understand the story the tumor is telling us, but we have reached the limit of what our brains and eye can do,” said Dr. Joseph Carlson, a professor and chief of the division of anatomic pathology at City of Hope and former associate director of surgical pathology at the USC Keck School of Medicine. “The information is hidden in there, but you need the right tool to unlock it.”
USC Professor Yan Liu and Dr. Joseph Carlson of City of Hope.
It turns out USC researchers like Yan Liu, a professor of computer science, electrical and computer engineering and biomedical sciences, may hold the key.
In a study, Liu, Carlson and their teams introduce a groundbreaking algorithm designed to predict how ovarian cancer tumor respond to standard treatments.
Ovarian cancer, one of the most serious malignancies affecting the female reproductive system, is diagnosed in approximately 21,000 people annually, with an estimated 14,000 deaths each year, according to the American Cancer Society. It’s also one of the most challenging to diagnose and manage, said Carlson, a specialist in gynecologic pathology.
“In part, this is because it is difficult to tease out the differences between ovarian cancer and other cancers,” said Carlson. “It’s also relatively rare, so we have less experience to work from.”
“It is difficult to tease out the differences between ovarian cancer and other cancers.” Dr. Joseph Carlson
Working with Carlson, Liu and her team developed a deep-learning system capable of accurately predicting treatment effectiveness on both a personalized level and a large scale. After training their model using a dataset of 288 patients, the researchers tested it in the context of ovarian cancer, in which roughly 20% of tumors persist after treatment.
“That means 20% of patients will be extremely ill and have no benefit,” said Carlson. “Their tumor will just continue to grow, and we can’t see why by looking in the microscope. If you knew ahead of time which patients would be resistant to treatment, you could tailor their treatment much more.”
The model successfully identified tumors that would respond to therapy, and those likely to resist standard treatment, which typically involves heavy-duty chemotherapy agents such as platinum and CarboTaxol. The model’s ability to integrate information from multiple sources—mirroring real-world diagnostic procedures—is key to its success, said Liu.
“We’re using data from different sources – blood samples, image data, and high-resolution digital scans of entire pathology slides,” said Liu.
“Each patient is different, and it’s crucial to build a personalized treatment recommendation system,” she added. “By adapting to the important features for each patient, we can provide explainability in terms of clinical variables and how they play a role across different stages of cancer.”
Beyond predicting treatment responses, Liu’s system generates “attention maps” that reveal what the AI is focuses on when making its predictions, shedding light on its decision-making process. For instance, is there something unique about these specific cells compared to all the other areas?
“This is one of the aspects that I like most as a pathologist,” said Carlson. “It’s a chance to catch flaws and errors in the system, but also for clinicians to learn from the AI system to see if we can use what the AI is showing us to then better understand why certain tumors are resistant or sensitive.”
The team also successfully transferred the model to kidney cancer dataset, hinting at the potential scope of the technology.
“As a pathologist, I’m like, ‘wow,’” said Carlson. “I am very hopeful that the way that we’re making decisions about a lot of things in healthcare is going to start to change and evolve and incorporate these types of methods. I do think this is the future.”
Published on December 11th, 2024
Last updated on December 18th, 2024
