Image Segmentation based Deep Learning for Biliary Tree Diagnosis

Abstract

Dilation of biliary tree can be an indicator of several diseases such as stones, tumors, benign strictures, and some cases cancer. This dilation can be due to many reasons such as gallstones, inflammation of the bile ducts, trauma, injury, severe liver damage. Automatic measurement of the biliary tree in magnetic resonance images (MRI) is helpful to assist hepatobiliary surgeons for minimally invasive surgery. In this paper, we proposed a model to segment biliary tree MRI images using a Fully Convolutional Neural (FCN) network. Based on the extracted area, seven features that include Entropy, standard deviation, RMS, kurtosis, skewness, Energy and maximum are computed. A database of images from King Hussein Medical Center (KHMC) is used in this work, containing 800 MRI images; 400 cases with normal biliary tree; and 400 images with dilated biliary tree labeled by surgeons. Once the features are extracted, four classifiers (Multi-Layer perceptron neural network, support vector machine, k-NN and decision tree) are applied to predict the status of patient in terms of biliary tree (normal or dilated). All classifiers show high accuracy in terms of Area Under Curve except support vector machine. The contributions of this work include introducing a fully convolutional network for biliary tree segmentation, additionally scientifically correlate the extracted features with the status of biliary tree (normal or dilated) that have not been previously investigated in the literature from MRI images for biliary tree status determinations.