Emerging research suggests than a radiomic AI model may significantly improve early detection of pancreatic ductal adenocarcinoma (PDA) on computed tomography (CT) scans.
For the study, recently published in Gut, researchers trained and evaluated the automated Radiomics-based Early Detection MODel (REDMOD), which combines AI-powered segmentation with 40 radiomic features. The multicenter training cohort was comprised of 156 pre-diagnostic PDA CT scans and 813 control CTs, according to the study authors. Subsequent independent testing of REDMOD was performed on a cohort comprised of 63 pre-diagnostic PDA CT scans and 430 controls.
In independent testing, the researchers found that REDMOD provided an 82 percent AUC, a 73 percent sensitivity rate and an 81.1 percent specificity rate for detecting stage 0 PDA.
Additionally, the study authors pointed out that the median time between CT detection of stage 0 PDA with REDMOD and histopathological diagnosis was 475 days.
“This temporal window holds profound significance as attaining such early detection would substantially augment the probability of cure and improved survival,” noted lead study author Sovanial Mukherjee, MD, who is affiliated with the Department of Radiology at the Mayo Clinic in Rochester, Minn., and colleagues.
In comparison to two reviewing abdominal radiologists with three years of post-residency experience, REDMOD had a 34.1 percent higher sensitivity rate (73 percent vs. 38.9 percent) at the median 475-day lead time and a nearly threefold higher sensitivity (68 percent vs. 23 percent) at a greater than two-year lead time, according to the researchers.
Three Key Takeaways
• Promising early detection performance on CT. The REDMOD radiomics AI model achieved strong diagnostic accuracy for stage 0 pancreatic ductal adenocarcinoma (AUC 82 percent, sensitivity 73 percent, specificity 81.1 percent) in independent testing, supporting its potential role in identifying preclinical disease.
• Substantial lead time before clinical diagnosis. REDMOD detected pancreatic cancer a median of 475 days earlier than histopathologic diagnosis, highlighting a clinically meaningful window for earlier intervention and potential survival benefit.
• Outperforms radiologists and leverages subtle imaging features: The model demonstrated markedly higher sensitivity than early-career radiologists (up to ~3× higher at >2-year lead times), with performance driven largely by filtered radiomic features that capture subtle, diffuse textural changes as opposed to discrete lesions.
The researchers noted that filtered radiomic features comprised 90 percent of the radiomic features employed in REDMOD, and demonstrated an 8 percent higher AUC in comparison to unfiltered radiomic features in an ablation analysis.
“This suggests that the earliest manifestations of PDA are subtle multiscale textural disruptions and alterations in local intensity gradients, which are more effectively captured by these advanced feature engineering techniques,” added Mukherjee and colleagues.
(Editor’s note: For related content, see “Can a New MRI-Based Risk Stratification Model Bolster Survival Prediction with Pancreatic Ductal Adenocarcinoma?,” “CT-Based Risk Scoring System Outperforms AJCC TNM Staging for Predicting Recurrence of Pancreatic Ductal Adenocarcinoma” and “Study Shows Promise of Emerging Radiotherapeutic Agent for Treatment of Gastroenteropancreatic Neuroendocrine Tumors.”)
In regard to study limitations, the authors acknowledged a lack of comparison of AI results across different ethnic and racial groups. Noting the lack of a discernible focal lesion for ground truth purposes in training an algorithm for localized detection at an early occult stage, the researchers said the AI model was designed to assess the entire pancreatic gland.
