Development of an artificial intelligence model to identify duodenal polyps in patients with Familial Adenomatous Polyposis (FAP)
Endoscopy · 01.04.2024 · 56 (S 02 S 02) · S121
DOI: 10.1055/s-0044-1782955 ↗Kurz gefasst
Fachartikel in der Zeitschrift Endoscopy, erschienen am 01.04.2024. Verfasst von D. Schupack und 6 weiteren Autor:innen.
Abstract (Originalsprache)
Aims
To develop an convolutional neural network model to identify duodenal polyps in patients with familial adenomatous polyposis.
Methods
We utilized a database of FAP patients who underwent upper GI tract endoscopic surveillance from 1989-2023 at our institution, to obtain high quality images (lacking significant artifact – blood, instruments, blurring) containing duodenal polyps. Ampullary images were not included to avoid bias. Annotation of polyps were completed using LabelStudio. We then employed nnUNet framework, a self-configuring method for deep learning-based biomedical image segmentation, to automate segmentation of duodenal polyps. Our dataset of 870 images was divided into training, validation, and test sets (80/10/10% split). The model was trained for 500 iterations. The primary metric for evaluating model performance was Dice coefficient, a statistical score between 0 (no agreement) and 1 (perfect agreement) used to gauge similarity between the model predictions and ground truth annotations. Using manual counting, we also compared the ability of the model to identify polyps on the set of 87 test images.
Results
Per manual count, mean polyps/image was 4.2 (range 1-27, median 2) and by prediction model mean was 4.0 (range 1-19, median 3). On 63/87 images (72.4%) the model missed 0 polyps, on 73/87 (83.9%)≤1 polyp was missed, and on 84/87 (96.6%)≤4 polyps were missed. Overall, mean missed polyps per image was 0.8 (range 0-14) and 297/367 (80.1%) duodenal polyps were identified by the model. In the 3 images with>4 missed polyps, mean polyp number was 22.3 (range 16-27). Falsely identified polyps occurred in 34/87 (39.0%) with a mean of 0.6 per image (range 0-4). There was complete match (no missed polyps, no false positive polyps) in 39/87 (44.8%). On our test set, the Dice coefficient was 0.7.
Conclusions
We successfully developed a model to identify duodenal polyps in FAP patients with reasonable accuracy. While the Dice coefficient of 0.7 is modest in comparison to models created for colon polyp detection, the differing aspects of anatomy and background mucosa in the duodenum makes this a challenging location for both human and computer modeling detection. Beyond this, the simple rate of polyp detection, likely a better marker of achieving our goal, was>80%, with relatively low rate of false polyps (
Zitieren
APA
Schupack, D., Sood, S., Shalini, S., Fetzer, J., Leggett, C., Boardman, L., & Coelho-Prabhu, N. (2024). Development of an artificial intelligence model to identify duodenal polyps in patients with Familial Adenomatous Polyposis (FAP). Endoscopy, 56(S 02 S 02), S121. https://doi.org/10.1055/s-0044-1782955
Vancouver
Schupack D, Sood S, Shalini S, Fetzer J, Leggett C, Boardman L, et al. Development of an artificial intelligence model to identify duodenal polyps in patients with Familial Adenomatous Polyposis (FAP). Endoscopy. 2024;56(S 02 S 02):S121. doi:10.1055/s-0044-1782955
BibTeX
@article{Schupack2024,
author = {Schupack, D. and Sood, S. and Shalini, S. and Fetzer, J. and Leggett, C. and Boardman, L. and Coelho-Prabhu, N.},
title = {Development of an artificial intelligence model to identify duodenal polyps in patients with Familial Adenomatous Polyposis (FAP)},
journal = {Endoscopy},
year = {2024},
volume = {56},
number = {S 02 S 02},
pages = {S121},
doi = {10.1055/s-0044-1782955}
}
EUVIMED