Background and aims:
EUS-guided needle-based confocal laser endomicroscopy (EUS-nCLE) can effectively differentiate high-grade dysplasia/adenocarcinoma (HGD-Ca) in branch duct (BD) intraductal papillary mucinous neoplasms (BD-IPMNs). However, this process requires manual image review and interpretation that is susceptible to subjectivity. Our objective is to present a working nCLE-AI algorithm capable of automatically editing full-length, unedited nCLE videos and risk-stratifying BD-IPMNs.

Methods:
Participants with a reference histopathological diagnosis of BD-IPMNs were selected from two prospective studies evaluating EUS-nCLE: (i) INDEX, 2015-2018 (NCT02516488), and (ii) CLIMB, an ongoing multi-center study (2018-present) (NCT03492151).

We developed two CNN-based algorithms (Figure 1): the first to effectively edit nCLE videos to high-yield frames with pathognomonic papillary structures, using YOLO-v8 and Inception V3 models; the second to automatically extract nCLE features for BD-IPMN risk stratification, using another Inception V3 model. Five rounds of experiments (with different training-validation-test data splits) were conducted to ensure that every participant was included in the test set at least once. The AI algorithm's ability to detect HGD-Ca was compared to histopathology, and diagnostic parameters were calculated and compared with the revised 2017 Fukuoka High Risk (HR) criteria.

Results:
EUS-nCLE imaging data from a total of 66 participants (mean age 68 ± 9 years, mean cyst diameter 35 ± 10 mm) were included. 41% of BD-IPMNs were HGD-Ca on surgical histopathology. The AI algorithms successfully edited full-length nCLE videos from the test set (n=66 across five rounds) and detected HGD-Ca with a sensitivity, specificity, and diagnostic accuracy of 78% (95% CI 59-89%), 72% (95% CI 56-84%), and 74% (95% CI 63-83%), respectively (Table 1).

In contrast, the 2017 Fukuoka HR performance yielded lower sensitivity of 44% (95% CI 28-63%), higher specificity of 95% (95% CI 83-99%), and comparable diagnostic accuracy of 74% (95% CI 63-83%).

Combining diagnostic parameters of the EUS-nCLE AI algorithm with the 2017 Fukuoka HR criteria (Table 1), the overall performance was 85% sensitive (95% CI 68-94%), 69% specific (95% CI 54-81%), and 76% accurate (95% CI 64-85%).

Conclusion:
We present a CNN-AI algorithm for nCLE video editing and risk stratification of BD-IPMNs. Our model matches established guidelines in diagnostic performance. Its key strength is impartiality, ensuring reproducibility across users. AI's nature enables instantaneous results, reducing nCLE diagnosis time. An added advantage is adaptability for future enhancements, including integrating demographics, clinical, and molecular data. For further enhancement and automation, extensive databases and prospective validation are recognized as necessary.

Zarine Shah

Appearances