ARTIFICIAL INTELLIGENCE ENHANCED ADEQUACY TESTING FOR EUS FNA SAMPLES FROM SOLID PANCREATIC LESIONS: PROSPECTIVE BLINDED TRIAL

Introduction:
Rapid on-site evaluation (ROSE) can provide immediate assurance of sample adequacy; however, significant manpower and logistic challenges have affected overall utilization of ROSE during endoscopic ultrasound guided fine needle aspiration and biopsy (EUS-FNA/B). We report results from a clinical trial evaluating the performance of an artificial intelligence ROSE software (ROSEAId) in comparison to in-person ROSE performed by a cytopathologist.
Methods:
Patients with solid pancreatic masses undergoing EUS-FNA procedure were enrolled in an IRB approved single blinded clinical trial. Stained pathology slide after each EUS-FNA needle pass, was evaluated by gastroenterologist with help of ROSEAId and data regarding adequacy, benign vs malignant and potential diagnosis by ROSEAId was documented. The slide was then evaluated by a blinded cytopathologist in the room. AI data was compared to the final cytopathology results. Clinical decisions were based upon the interpretation of the cytopathologists.
Results:
Total 50 patients with initial imaging concerning for pancreatic solid lesions undergoing EUS were consented for the study. 11/50 patients were excluded from the final analysis (4 peripancreatic masses, 5 cystic lesions and 2 no definitive pancreatic mass). Final cohort included 39 patients: 32 with final diagnosis of adenocarcinoma, 2 with neuroendocrine tumor, and 5 benign cases (2 Granulomatous pancreatitis, 1 Rosai Dorfman Syndrome, 1 chronic pancreatitis, 1 inflammatory mass). In all 39 cases, gastroenterologists were able to use DWSS and get image interpretations using ROSEAId.
On Site Adequacy: 97% (38/39) patients’ sample were deemed adequate by ROSEAId’s compared to 92% (36/39) by cytopathologists.
Malignant Lesions: ROSEAId accurately diagnosed 34/34 (100%) cases, however it misclassified 3 malignant lesions. One case of neuroendocrine tumor was misclassified as adenocarcinoma and 2 cases of carcinoma were misclassified as adenocarcinoma.
Benign Lesions: ROSEAId accurately identified 2/5 (40%) of benign cases. 2 cases of granulomatous pancreatitis and 1 case of chronic pancreatitis were misclassified as adenocarcinoma by ROSEAId. Both cases of granulomatous pancreatitis were classified as likely malignancy by on site review of pathologists as well. ROSEAId and Cytopathlogists onsite assessment was similar on 4/5 (80%) cases.
Conclusions: Overall, ROSEAId was easy to use and incorporate in clinical practice. In the initial analysis so far, we found ROSEAId to be accurate in determining onsite adequacy of the sample.