POLYP FINGERPRINT: AUTOMATIC RECOGNITION OF UNIQUE FEATURES TO UNIVOCALLY IDENTIFY COLORECTAL POLYPS

Aims:

Following human recognition by fingerprints, we propose to study the potential of computer systems in the definition and recognition of unique characteristics for each colorectal polyp. Polyp fingerprint can be used in two types of applications: assistance in polyp detection to ensure that a polyp observed during insertion is recognized during withdrawal and assistance in the in-vivo prediction of the histology of those polyps sharing a similar appearance with one with known histology.

Methods:

Our system uses a color descriptor to characterize the image and applies Bag of Words technique to build a vocabulary univocally describing each image. To test the methodology, we used 225 images from 76 polyps acquired during routinely explorations at Hospital Clinic of Barcelona using high definition OLYMPUS endoscopes. At least two images showing different views from the same polyp were used in the experiment. The automatic system provides for each image the closest match within the dataset.

Results:

The distribution of polyps according to Paris classification was: 40 of type 0-Is (118 images), 31 of type 0-IIa (11 images) and 5 of type 0-Ip (96 images). Mean polyp size was of 11,60 mm. 61 out of 76 polyps were adenomas (80.26%, 173 images). In our experiment, 207 images (92%) matched another image of the same polyp. For those polyps with only two images, the system provided an accurate match in 31/33 cases (93.94%). In the subset of images where the polyp was represented with more than two images, the system provided an accurate match in 42/43 cases (97.67%). Furthermore, in this subset the system provided correct matches for all images of the same polyp in 31/43 cases (72.09%).

Conclusions:

A computational system can accurately recognize as a unique lesion a polyp observed in different views by describing the endoluminal scene using a color descriptor.