HISINVIA: A HYBRID SOLUTION FOR COLONIC POLYP HISTOLOGY PREDICTION IN WHITE LIGHT COLONOSCOPY IMAGES COMBINING ARTIFICIAL INTELLIGENCE AND CLINICAL INFORMATION

Aims In-vivo histology prediction is the cornerstone to improve the cost-efficiency of colonoscopy procedures. Artificial Intelligence (AI), specifically via Deep learning (DL) systems, can help physicians during colonoscopy in this task. However, its efficiency has not yet reached the levels of performance necessary to be used in the exploration room. In order to improve that, we propose a hybrid approach (HISINVIA) that combines DL methods with polyps characteristics indicated by doctors.

Methods HISINVIA combines features extracted from the polyp region in 1346 original white light HD colonoscopy images from 501 different polyps using a DL architecture (ResNet50) with three different features indicated by the clinicians: size (in mm), location and morphology according to Paris classification. Polyp region in each image was delineated by clinicians using GTCreator software. The system provided as output the probability of being an adenoma vs non-adenoma.

Results 926 images (69%) contained an adenomatous polyp whereas 419 (31%) showed a non-adenomatous polyp. 579 (43%) were < 5 mm, 331 (25%) 6–10 mm and 436 (32%) > 10 mm. Regarding the location, 598 (45%) were in the rectum-sigma and 167 (12%) in the right colon. Images were distributed into train (70%, 940 images), validation (20%, 269 images) and test (10%, 137 images) sets. Deep learning information was able to correctly identify 78/94 adenomas and 17/43 non-adenomas. HISINVIA was able to correctly identify 80/94 adenomas and 32/43 non-adenomas. Moreover, the processing time was very low (50 ms to process an image). Complete performance metrics are shown in the table.

Tab. 1

	Sens	Spec	PPV	NPV
Deep learning	83%	40%	75%	52%
HISINVIA	85%	74%	88%	70%

Conclusions HISINVIA improves performance obtained by pure AI-based solutions, showing its potential to be used in the exploration room.