Auto-detection Of Lesions On Endoscopic Images Of The Esophagus

Reflux diseases and esophageal lesions are common types of illness attacking the human esophagus.Diagnosis is often done by computed tomography(CT)and magnetic resonance(MR)imaging.Endoscopic images can also be collected and processed separately using various image processing teachings,such a pattern recognition and segmentation.The later is a widely used method in computer vision and images processing.It allows us to partition a set of data,based on certain characteristics or features on the sample data.The application in medical image analysis,especially for abnormal tissues or lesions detection,has evolved very rapidly in the past few decades.The accuracy and efficiency of the technique used to detect lesions is in direct correlation with the type and the amount of lesions that can be identified using such techniques.This thesis proposes a new way of detecting esophageal lesions,with focus on granular cell tumors and Barrett’s esophagus detection.Endoscopic image of the esophagus are collected and processed as 2 D image in the context of this thesis.We leverage both the color and texture properties of each image to uncover eventual lesions.For granular cell tumors,lesions appear in granular shapes or blobs of various size and depth.But these lesions mostly fall in a certain size range which can be exploited to single them out.Therefore,in our project,blobs finding methods are first applied to the image to find all the large blobs of maximum estimated size,then find small blobs with a minimum defined size.Once blobs are all discriminated by size,a final lesion confirmation algorithm is involved to determine which of the blobs among those matching the size boundary,are actually lesions.As for Barrett’s esophagus detection,we use the k-means clustering method to segment the image based on given features.Two main types of features are extracted: The color and texture features.Because all lesions in the case of Barrett’s esophagus appear with a color and texture different from normal tissues.The color features are generated from the HSV color properties of the image,while Gabor filter banks are used for the texture features extraction.Some spacial features are also used to maintain adjacency.The proposed methods proved to be very efficient and accurate,even for image from various data sources.