Integrating artificial intelligence with endoscopic ultrasound for the differential diagnosis of pancreatic solid neoplasms

Aims Solid pancreatic lesion incidence has risen in recent years with Pancreatic ductal adenocarcinoma (PDAC) and pancreatic neuroendocrine neoplasms (PanNENs) representing the two most prevalent, each requiring distinct treatments. Endoscopic ultrasonography (EUS) has proven to be a highly reliable technique for diagnosing pancreatic solid lesions, however, its diagnostic accuracy is constrained by operator experience and moderate interobserver agreement. The increasing popularity of artificial intelligence (AI) in medicine has led to the development of AI-based models capable of processing extensive data volumes, with promising results in various medical applications. This study aimed to assess the effectiveness of our AI-based model in analyzing EUS images of PDAC and PanNENs and distinguishing between the two types of lesions. The secondary aim is to assess the model's capability in lesion segmentation.

Methods Consecutive patients who underwent EUS at our center and received a pathological diagnosis of PDAC or PanNENs through biopsy or surgery were included in this retrospective study. One image per patient, containing the lesion, was selected, anonymized, and manually segmented by an experienced gastroenterologist. Images of low quality, with artifacts or obtained after contrast administration, were excluded. The image dataset was then augmented through data augmentation protocols and divided into training, validation, and test cohorts with a ratio of 60:20:20. Using these cohorts, we developed an AI model employing a deep-learning architecture to classify and segment the lesions.

Results The study included 307 patients (201 with PDAC and 106 with PanNENs), predominantly male (54.4%), with a median age of 66 years at diagnosis. Two AI models, Model 1 and Model 2, were tested across 5-folds cross-validation for classification and segmentation tasks. Model 1 demonstrated mean average precision, Receiver Operating Characteristic Area Under the Curve (ROC AUC), and balanced accuracy at 87.48%, 80.12%, and 71.00%, with specificity and sensitivity at 60.00% and 82.00%, and an Intersection over Union (IOU) of 53.65% in segmentation. Model 2, focusing on the classification task, showed improved performance with mean average precision, ROC AUC, and balanced accuracy at 88.90%, 82.00%, and 73.63%, and specificity and sensitivity at 64.76% and 82.50%, respectively. [1] [2] [3] [4] [5] [6]

Conclusions The AI model we developed can differentiate between PDAC and PanNENs; however, further validation with larger datasets is required.

Conflicts of interest

Authors do not have any conflict of interest to disclose.

• References

• 1 Huang B, Huang H, Zhang S. et al. Artificial intelligence in pancreatic cancer. Theranostics 2022; 12 (16) 6931-6954
• 2 Kaul V, Enslin S, Gross SA.. History of artificial intelligence in medicine. Gastrointest Endosc 2020; 92 (04) 807-812
• 3 Gardner TB, Gordon SR.. Interobserver agreement for pancreatic endoscopic ultrasonography determined by same day back-to-back examinations. J Clin Gastroenterol 2011; 45 (06) 542-545
• 4 Zilli A, Arcidiacono PG, Conte D, Massironi S.. Clinical impact of endoscopic ultrasonography on the management of neuroendocrine tumors: lights and shadows. Dig Liver Dis 2018; 50 (01) 6-14
• 5 Kitano M, Yoshida T, Itonaga M, Tamura T, Hatamaru K, Yamashita Y.. Impact of endoscopic ultrasonography on diagnosis of pancreatic cancer. J Gastroenterol 2019; 54 (01) 19-32
• 6 Siegel RL, Miller KD, Jemal A.. Cancer statistics, 2020. CA Cancer J Clin 2020; 70 (01) 7-30

Publication History

Article published online:
15 April 2024