Segmentation of Polyps in Gastrointestinal Tract Images

Sabrina Nasrin (1) , Javaneh Alavi (1) , Pamila Viswanathan (1)

1 University of Alberta

Detecting abnormal tissues can be overlooked during body screening procedures including endoscopy, bronchoscopy, and colonoscopy. Colonoscopy is a routine screening procedure that can examine inside of the large intestine. However, observants might not be able to detect anomalies at initial phase. Therefore, a precise method is needed to detect the abnormalities. In this paper, we have implemented three different convolutional neural networks to segment polyps in gastrointestinal tract images. First, UNet which consist of two parts contraction and expansion for segmenting medical images. In this model data augmentation is performed with elastic deformations to yield accurate results with very few annotated images.Then, we implemented TriUnet which consists of three UNet models. The last model DivergentNets is an ensemble of five segmentation models named as TriUnet, Unetplusplus, FPN, DeeplabV3 and DeeplabV3plus. We have also tested images by using color correction, image pyramid and specularity removal. Our results suggest that when we combine different segmentation models as DivergentNets, it produces better results than UNet and TriUnet.

Domaines: Intelligence artificielle [cs.AI]

Sabrina Nasrin, Javaneh Alavi, Pamila Viswanathan. Segmentation of Polyps in Gastrointestinal Tract Images. [Research Report] University of Alberta. 2021. ⟨hal-03501140⟩