AI model highly accurate in differentiating pancreatic cystic, solid lesions on ultrasound
Click Here to Manage Email Alerts
Key takeaways:
- AI model demonstrated 94% accuracy in discerning cystic pancreatic focal lesions from solid lesions in endoscopic ultrasound.
- Model “has potential to transform” pancreatic focal lesion exams.
PHILADELPHIA — An AI model achieved high accuracy in detecting and distinguishing between cystic and solid pancreatic focal lesions during endoscopic ultrasound procedures, according to data presented at the ACG Annual Scientific Meeting.
“Pancreatic focal lesions, either cystic or solid, are commonly found in imaging exams,” Miguel Mascarenhas, MD, PhD, from the University Hospital Center of São João in Porto, Portugal, told Healio. “Endoscopic ultrasound is fundamental in the management of patients with pancreatic focal lesions, assuring single procedure lesion analysis and tissue sampling. However, it is an operator-dependent procedure with high interobserver variability.”
He noted that “the differential diagnosis of pancreatic solid lesions includes pancreatic adenocarcinoma, with poor prognosis due to late-stage disease at diagnosis. On the other hand, the management of cystic lesions is highly dependent on lesion type, given the almost exclusive association of malignancy risk with cystic lesions with a mucinous phenotype.”
To achieve “more accurate and earlier diagnosis” for these patients, Mascarenhas and colleagues developed a convolutional neural network (CNN) based on 378 endoscopic ultrasound exams from four international reference centers in Brazil, Portugal, Spain and the U.S.
With the intention to differentiate between pancreatic cystic neoplasms (mucinous and non-mucinous lesions) and solid pancreatic lesions (adenocarcinoma/neuroendocrine tumor), the CNN included more than 126,000 endoscopic ultrasound images, consisting of 19,528 mucinous, 8,175 non-mucinous, 64,286 adenocarcinoma, and 29,153 neuroendocrine tumor images as well as 4,858 normal pancreatic images.
According to study results, the CNN demonstrated a 99.1% accuracy for identifying normal pancreatic tissue, with 99% accuracy for mucinous pancreatic cystic neoplasms and 99.8% for non-mucinous neoplasms. The model was able to distinguish pancreatic adenocarcinoma and pancreatic neuroendocrine tumor with 94% accuracy, with a sensitivity of 98.7% for adenocarcinoma and 83.7% for neuroendocrine tumors.
“Our group developed the first worldwide AI model capable of detection and differentiation of pancreatic cystic lesions and pancreatic solid lesions,” Mascarenhas told Healio. “The development of AI models is of uttermost importance for augmenting endoscopic ultrasound diagnostic accuracy, and the availableness of a technology that is efficient in detecting pleomorphic pancreatic lesions is a proof of methodological development and robustness.”
Although there were significant limitations to the study, namely its retrospective design and that it used still images as opposed to video, Mascarenhas noted that “future studies will focus in multicentric real-time validation of the technology during endoscopic ultrasound procedures.”
“Endoscopic ultrasound is essential in the management of patient with pancreatic focal lesions,” he told Healio. “AI has the potential to transform this exam through more accurate diagnosis, better detection of smaller lesions and guidance of biopsies, assuring better patient management.”
Collapse
Mascarenhas M, et al. Artificial intelligence-driven endoscopic ultrasound for detection and differentiation of pleomorphic pancreatic lesions: A transatlantic study. Presented at: ACG Annual Scientific Meeting; Oct. 25-30, 2024; Philadelphia (hybrid meeting).
Read more about
Click Here to Manage Email Alerts