Colon tumor localization using three input variants to Faster Region‐based Convolutional Neural Network and lazy snapping

Abstract

AbstractAutomated polyp localization in colon endoscopy images helps minimize human errors in localizing polyps. In our method, we use Faster Region‐based Convolutional Neural Network (R‐CNN) on Resnet 50 network to form a tight bounding box around the polyp. The bounding box is then used as input for the lazy snapping technique to determine polyps correctly. Three input variants—RGB images, histogram equalized images, and luminance images—are fed to the network. The output obtained from each variant is combined to form the final result. We have used the CVC‐Clinical DB database, which has 612 images with 672 polyp instances for our study. Thirteen different combinations for obtaining the result are studied, and the best among them is identified. The result is evaluated for all combinations and against a state‐of‐the‐art method for precision, recall, and F‐measure. The proposed model achieves a precision of 80.51% and a recall value of 80.33%.