DEEP LEARNING AND DIGITAL SINGLE-OPERATOR CHOLANGIOSCOPY (DSOC): AUTOMATIC DIAGNOSIS OF MALIGNANCY STATUS AND MORPHOLOGICAL CHARACTERIZATION OF BILIARY STRICTURES

Aims Patients with indeterminate biliary strictures (BS) constitute a significant diagnostic challenge. DSOC has enabled morphologic characterization and guided biopsies. However, the diagnostic yield of DSOC remains suboptimal, and the characterization of these lesions has significant interobserver variability.

With this work, we intend to develop a Convolutional Neural Network (CNN), for detection of malignant BS in DSOC images and identification of three morphologic features: nodules (NN), papillary projections (PP), and tumor vessels (TV).

Methods We developed and validated a CNN based on DSOC images. Each image was labelled as a normal/benign finding or a malignant lesion if definite histologic evidence of malignancy was available.

The CNN was also trained to detect morphologic features associated with biliary malignancy: NN, PP, and TV. The performance of the CNN was then tested.

Results We used 23595 images from 125 patients (20719 of malignant BS and 2876 of normal or benign findings).

The model presented a sensitivity of 98.9%, specificity of 97.7%, and overall accuracy of 98.7%.

Additionally, the model comprised 2876 images of NN, 1675 images showing PP, and 4153 images of TV. The accuracy for the detection of these features was, respectively, 96.9%, 96.1% and 91.5%.

Conclusions Our group developed a pioneer combined CNN for the simultaneous detection of malignant BS and identification of morphologic features associated with malignancy. Applying AI models to DSOC may increase its diagnostic yield for patients with indeterminate BS. Furthermore, accurate real-time identification of those features may help to guide biopsies, thus increasing their reliability and the diagnostic yield of DSOC.

Publication History

Article published online:
14 April 2022