From challenge to change: AI's leap in early pancreatic cancer identification

Nov. 30, 2023

Mayo Clinic researchers have introduced an AI technology that has potential to revolutionize early pancreatic cancer detection. At the center of their technology is an automated AI model that accurately detected pancreas cancer in diagnostic CT scans and even flagged hidden cancers in pre-diagnostic scans at around 475 days before their eventual clinical diagnosis. These groundbreaking findings were recently published in the journal Gastroenterology.

"The ability to identify cancer using AI-augmented standard-of-care imaging in seemingly normal pancreases at the pre-diagnostic stage is a major advancement in the field."

— Ajit H. Goenka, M.D.

Pancreatic cancer is projected to become the second leading cause of cancer deaths in the U.S. by 2030. Early detection remains a challenge. Most patients are diagnosed at advanced stages where treatment options are limited, and nearly 70% don't survive past the first year. Standard imaging often fails to identify up to 40% of small pancreatic cancers — including in high-risk individuals actively being screened — due to either subtle imaging features, inadequate attention to the pancreas or technical inadequacies. This creates a crucial "last-mile" obstacle in early detection efforts.

"Early detection of pancreatic cancer could substantially improve the chance of curative surgery for many patients," says Ajit H. Goenka, M.D., a nuclear medicine specialist and radiologist at Mayo Clinic Comprehensive Cancer Center in Rochester, Minnesota, and the study's senior author. "The ability to identify cancer using AI-augmented standard-of-care imaging in seemingly normal pancreases at the pre-diagnostic stage is a major advancement in the field."

Data-driven excellence: Advancing CT diagnostics with big data and modern programming

Dr. Goenka's team trained the automated AI model on an extensive dataset of standard-of-care imaging, including over 3,000 patient CT scans, making it one of the largest and most diverse datasets used for such research. The majority of CTs (64%) were from external institutions, which underscores the diversity of the training dataset. The model had a mean accuracy of 92%, correctly classifying 88% of CT scans with cancer and 94% of control CT scans. The area under the receiver operating characteristic (AUROC) curve was 0.97, with a sensitivity of 88% and a specificity of 95%.

Even though the model was trained on CTs with larger tumors, the model detected pancreatic cancer in pre-diagnostic CTs taken between 3 to 36 months (median 475 days) before the cancer's clinical diagnosis. In these scenarios, the model reported an accuracy of 84%, an AUROC curve of 0.91, a sensitivity of 75% and a specificity of 90%. These findings highlight the model's potential in early pancreatic cancer detection and intervention.

The model demonstrated reliability and consistent accuracy across diverse patient demographics, scanner technology and imaging protocols. Such robustness is essential for the model's applicability in real-world medical scenarios.

Demystifying AI's choices: A step toward transparent decision-making

To combat concerns about AI acting as a "black box," Dr. Goenka's team shed light on its decision-making process. Recognizing the critical importance of transparency, trustworthiness and quality in AI's clinical adoption, the team highlighted its efforts to make AI operations more understandable.

The model is fully autonomous, which makes it apt for incorporation into established imaging protocols, negating the need for alterations to the current imaging workflow. With full automation, it has the potential to centrally analyze imaging data from multiple facilities without manual intervention.

The future of oncology: AI-augmented screening for early detection and risk stratification

The AI model can function as an adjunct reviewer to expert physicians, offering an enhanced scrutiny level by pinpointing subtle lesions or initial indicators of early cancer. Given its proficiency in forecasting subsequent cancer in pre-diagnostic scenarios, the model can enhance risk stratification for high-risk individuals, directing the frequency of screenings or the consideration of more-advanced diagnostic techniques.

Dr. Goenka's team is preparing for prospective trials, aiming to pioneer enhanced early screening methods for those most at risk of sporadic pancreatic cancer. This next exploration will be instrumental in fine-tuning and magnifying the real-world impact of this approach.

"The realm of AI's potential is vast and untapped," says Dr. Goenka. "Imagine a future where AI seamlessly integrates with advanced imaging and novel biomarkers, redefining our understanding of the disease biology and advancing the goal of early cancer detection."

For more information

Korfiatis P, et al. Automated artificial intelligence model trained on a large data set can detect pancreas cancer on diagnostic computed tomography scans as well as visually occult preinvasive cancer on prediagnostic computed tomography scans. Gastroenterology. 2023;165:1533.

Refer a patient to Mayo Clinic.