Endoscopy 2019; 51(04): S4-S5
DOI: 10.1055/s-0039-1681182
ESGE Days 2019 oral presentations
Friday, April 5, 2019 08:30 – 10:30: Artificial intelligence Club A
Georg Thieme Verlag KG Stuttgart · New York

BLI AND LCI IMPROVE POLYP DETECTION RATE AND DELINEATION ACCURACY FOR DEEP LEARNING NETWORKS

T Eelbode
1   KU Leuven, Medical Imaging Research Center, PSI, Leuven, Belgium
,
C Hassan
2   Nuovo Regina Margherita Hospital, Gastroenterology, Rome, Italy
,
I Demedts
3   KU Leuven, Department of Gastroenterology and Hepatology, Leuven, Belgium
,
P Roelandt
3   KU Leuven, Department of Gastroenterology and Hepatology, Leuven, Belgium
,
E Coron
4   Centre Hospitalier Universitaire Hotel Dieu, Hepatogastroenterology, Nantes, France
,
P Bhandari
5   Portsmouth University Hospital, Solent Centre for Digestive Diseases, Portsmouth, United Kingdom
,
H Neumann
6   University Medical Center Mainz, First Medical Department, Mainz, Germany
,
O Pech
7   Krankenhaus Barmherzige Brüder Regensburg, Department of Gastroenterology and Interventional Endoscopy, Regensburg, Germany
,
A Repici
8   Humanitas University, Digestive Endoscopy Unit, Milan, Italy
,
F Maes
1   KU Leuven, Medical Imaging Research Center, PSI, Leuven, Belgium
,
R Bisschops
3   KU Leuven, Department of Gastroenterology and Hepatology, Leuven, Belgium
› Author Affiliations
Further Information

Publication History

Publication Date:
18 March 2019 (online)

Also available at
 

Aims:

Studies have suggested that polyp detection rates can be improved by using other modalities than white-light imaging (WLI) such as linked-color imaging (LCI) from Fujifilm. Our aim is to evaluate the influence of the modality on polyp detection rate and delineation accuracy of an artificial intelligence (AI) system.

Methods:

Colonoscopy videos from 120 patients are included with a total of 280 polyps. Shorter video clips containing the first apparition of each polyp are extracted and for each clip, a few frames are annotated by experts. These 758 manual annotations are automatically propagated over the entire clip. The resulting, much larger annotated dataset of 40887 images is then used to train a recurrent convolutional neural network (CNN).

Frame-level sensitivity and specificity are reported for evaluation of the detection power of the network. For delineation accuracy, the dice score is used which is a measure for the amount of overlap between a delineation map and its ground truth. The analysis is done for WLI, BLI (blue light imaging) and LCI.

Results:

Table 1 shows that BLI significantly improves sensitivity, specificity and dice score. Similarly, LCI increases detection performance to a lesser extent, however the LCI Dice score decreases significantly compared to WLI. Pairwise t-tests show that all differences are significant with a p value < 0,00001 (significance level of 0,05).

Tab. 1:

Sample size, sensitivity, specificity and dice score (mean and stdev) for the three different modalities in the test set.

n

Sensitivity

Specificity

Dice score (mean & std)

WLI

151

0,81

0,76

0,69+-0,33

BLI

79

0,92

0,85

0,76+-0,28

LCI

58

0,85

0,82

0,63+-0,34

Conclusions:

The choice of modality has a significant impact on the detection and delineation performance of an AI system. We show that our network performs best for both tasks on BLI and that LCI has a superior detection, but inferior delineation power compared to WLI.