Artificial intelligence (AI) holds the promise of revolutionizing how we live, work, learn, and play. In medicine, AI is giving researchers and clinicians unique, immediate, and detailed insights into the human body and its biologic functions—insights that would take humans hours, days, weeks, or even years to uncover.
Brigham and Women’s Hospital and Dana-Farber Cancer Institute are at the forefront of harnessing the power of AI to change patients’ lives, as reflected in several recent research projects conducted by physician-scientists in Radiation Oncology and Radiology.
“We see an enormous potential for AI to democratize patient care and enable the tools needed to provide cutting-edge care globally,” says Naomi Lenane, chief information officer and vice president of Information Services, Dana-Farber Cancer Institute. “As we explore the bench-to-bedside implications of AI, we are doing so carefully, with full participation from multidisciplinary teams in research, clinical care, ethics, legal, and operations to ensure that we always lead with patient safety as our core goal.”
Predicting Pancreatic Cancer Before Symptoms Appear
In an effort to leverage AI to detect diseases earlier, Michael H. Rosenthal, MD, PhD, led a team that developed an AI algorithm that analyzes CT scans to differentiate skeletal muscle from fat tissue in patients with pancreatic cancer. They used the AI model to analyze 3,000 patient scans in hours rather than the months it would take to do so manually.
Their research, published in Nature Communications, shows that skeletal muscle wasting can be detected up to 18 months before a pancreatic cancer diagnosis and adipose tissue (fat) wasting can be detected about six months before. These findings indicate a potential signal or biomarker that could be used to find pancreatic cancer up to several years before it would otherwise be detected based on symptoms or other clinical findings.
“Early screening for skeletal muscle and fat wasting may be of particular benefit to patients with a high familial or genetic risk of developing pancreatic cancer, allowing clinicians to find tumors while they’re still easily resectable and patients have a higher potential for cure,” Dr. Rosenthal says. “Because we know early detection in high-risk families is very effective, we can suggest it also would be effective in the average risk population as well—and that they also may benefit from opportunistic screening.”
Optimizing Treatment for Head and Neck Cancer
Benjamin H. Kann, MD, and colleagues developed an AI platform that analyzes routine CT scans of patients with head and neck squamous cell carcinoma to diagnose sarcopenia (low muscle mass). The presence of this condition indicates a patient is at higher risk for treatment complications such as needing a long-term feeding tube or other supportive care.
The investigators’ work, published in JAMA Network Open, describes the development and validation of an AI model that automatically segments C3 skeletal muscle area to derive a skeletal muscle index for sarcopenia. Based on a review by an expert panel, the model made clinically acceptable assessments of muscle mass 96.2% of the time, completing scan assessments in about 0.15 seconds—a process that typically takes clinicians up to 10 minutes.
“Malnutrition and sarcopenia are serious problems in patients being treated for head and neck cancer,” Dr. Kann says. “Until development of this AI model, the only way to follow patients’ status in this area was with traditional nutritional screening, which tracks weight loss and body mass index [BMI]. Our tool allows clinicians to evaluate patient status using routine imaging instead. It is a more powerful predictor for knowing if a patient will become dependent on a feeding tube or be at risk for worse outcomes, including earlier death, than relying on weight or BMI alone.”
Avoiding Treatment Complications in Some HPV-related Oropharynx Cancer Cases
Dr. Kann also is involved in efforts to study the use of AI to help certain patients with human papilloma virus (HPV)-associated oropharynx cancer avoid trimodality therapy, which is associated with complications and poor quality-of-life outcomes. For these patients, assessing the presence of cancer cells beyond the lymph nodes, or extranodal extension (ENE), is critical for determining proper treatment but challenging to detect through imaging.
“HPV-related oropharynx cancer has become an epidemic, but it is associated with high cure rates and long survivorship,” Dr. Kann says. “At the same time, there are significant complications and toxicities associated with treatment. We can spare patients these complications by using a minimally invasive surgery called trans oral robotic surgery [TORS] instead of a combined course of chemo and radiation.”
The issue, according to Dr. Kann, is that patients with ENE are poor candidates for TORS. If ENE is discovered during TORS, which happens in about 50% of cases, the patient will still need chemotherapy and radiation.
Dr. Kann is corresponding author of a paper published in The Lancet Digital Health that retrospectively evaluated a CT-based deep learning algorithm that was found to widely outperform experts in predicting pathological ENE on a challenging cohort of patients. The study suggests the tool may be a promising way to detect ENE, assist in treatment decision-making, and guide optimal de-escalation strategies.
Taking the Prostate Contouring Process to a New Level
Martin T. King, MD, PhD, and colleagues are studying the use of AI to improve the process of prostate contouring for patients receiving transrectal ultrasound (TRUS)-based high-dose-rate prostate brachytherapy.
Their research, published in Medical Physics, suggests that an AI algorithm trained and validated using TRUS images and reference (clinical) contours generated high-quality results and good geometric accuracy with implanted needles. Furthermore, the AI contours were able to capture the complex deformations in prostate anatomy after needle insertion. Interestingly, independent observers preferred AI contours for most of the cases. The presence of AI contours also improved the geometric accuracy of resident contours.
As Dr. King notes, prostate contouring can be challenging and result in wide variability due to artifacts from implanted needles, bleeding, and calcification.
“Many clinicians, especially those early in their career, struggle with defining prostate volume correctly and consistently,” he says. “This AI tool may help to standardize care, provide an added level of quality assurance, and aid early-career physicians in implementing the procedure, all of which have important implications for prostate cancer treatment.”
Identifying and Monitoring Social Determinants of Health
Social determinants of health drive cancer disparities, and recent AI-based research shows that the prejudices and biases that exist in general society also are present in electronic health records (EHRs). Danielle S. Bitterman, MD, was part of a team that explored how natural language processing (NLP) can be used to automatically and empirically extract clinical documentation of social contexts and needs that may underlie disparities. She was corresponding author of a paper on the topic that was published in JCO Clinical Cancer Informatics.
“EHRs are an imperfect data repository and communications tool,” Dr. Bitterman says. “Our goal is to make EHRs more useful to providers and patients by transforming data that’s already in the EHR to enable more active patient monitoring and allow us to preemptively triage patients who need additional support. Essentially, we want to turn EHRs into a learning health system.”
Dr. Bitterman says it is critical to study potential human biases within EHRs so that AI models can be developed to recognize and exclude biases and other social contexts that drive inappropriate treatment decisions.
“NLP models are smart; they tend to learn the path of least resistance, similar to the way humans take mental shortcuts,” she says. “Therefore, if we typically give different treatments to people of different races and one of those treatments has better survival, the model will inappropriately use that demographic factor as a proxy for outcomes and learn that people of that race do better under that treatment.”
Physicians Express Caution Around AI Implementation
These physician-scientists share an excitement about AI’s potential as well as concerns about AI tools moving from the research lab to the clinic.
“AI will be fantastically beneficial across medicine, as long as we maintain thoughtful control over how it’s being used,” Dr. Rosenthal says. “AI systems can create the appearance of magic very readily by analyzing millions of data points in a mere fraction of the time it would take humans. Responsible and careful implementation is critical as we move from proof of concept into the clinical setting.”
Dr. Kann echoes the sentiment about AI’s ability to process and synthesize the vast number of data points collected across the spectrum of patient care. However, he notes that challenges in the ability to collect enough data to leverage AI to its fullest potential remain, especially given the siloed nature of the healthcare system. He stresses the need to tie AI’s power to clinically meaningful endpoints like improved survival and quality of life.
Dr. King adds that imaging is an ideal specialty in which to build trust in AI because images are easily double-checked by human observers to determine whether AI’s conclusions are believable. He cautions that while AI provides answers, it doesn’t necessarily reveal how it arrived at the answers, making it crucial to employ safeguards to ensure AI results reflect clinical scenarios.
“AI tools can have a positive impact on decision-making and clinical outcomes,” Dr. Bitterman concludes. “However, because it often runs in the ‘background,’ patients may not even be aware when AI is being used in their care, making it imperative that we hold these tools to the same type of Institutional Review Board and informed patient consent standards that we do for all medical devices and interventions.”