Subscribe to RSS
DOI: 10.1055/s-0040-1704526
AUTOMATED POLYP SIZE ESTIMATION WITH DEEP LEARNING REDUCES OVERESTIMATION BIAS
Publication History
Publication Date:
23 April 2020 (online)
Aims Polyp size is directly correlated with risk of future CRC and growth invasiveness. Despite its big impact, endoscopists typically provide a visual size estimation, but several studies have reported low accuracies. We aim to enable more accurate in-vivo polyp size measurements and ultimately reduce clinical mis-sizing by endoscopists. Therefore, we developed an AI system that can objectively infer polyp size based on a reference tool in the endoscopic image.
Methods We use a biopsy forceps as reference tool and apply two separate deep learning algorithms: (1) delineation of the polyp and (2) detection of two landmarks on the forceps. Since the exact dimensions of the forceps are known, we can compute the actual size of the polyp. For polyp delineation (1), we collected colonoscopy videos from 206 patients with 825 polyps for training the system. For the forceps detection (2), we collected videos from 41 patients with 69 polyps and extracted 289 frames containing a polyp and open forceps.
We report the trimmed average of the difference in size between the endoscopist or algorithm and the ground truth. The latter is defined by manually delineating the polyp and the two landmarks and is validated in a colon phantom.
Results The size estimation can detect the polyp and forceps in 71% of the test images. The trimmed average difference is + 0,52 mm (SD 1,78 mm) and + 1,40 mm (SD 1,82 mm) between the ground truth and predicted size by the algorithm or endoscopist respectively. Our algorithm thus leads to a decrease in overestimation bias by 63% (p-value < 0.1).
Conclusions We show that automated delineation of the polyp and forceps detection helps in estimating polyp size and significantly reduces the endoscopists’ estimation error. This can lead to better surveillance interval decisions, but we will evaluate this in a larger patient cohort.
#