Segmentation and polyp detection in virtual colonoscopy : a complete system for computer aided diagnosis

Abstract

El cancer colorectal es una de las mayores causas de muerte por cancer en el mundo. La deteccion temprana de polipos es fundamental para su tratamiento, permitiendo alcanzar tasas del 90% de curabilidad. La tecnica habitual para la deteccion de polipos, debido a su elevada performance, es la colonoscopia optica (tecnica invasiva y extremadamente cara). A mediados de los '90 surge la tecnica denominada colonoscopia virtual. Esta tecnica consiste en la reconstruccion 3D del colon a partir de cortes de tomografia computada. Es por ende una tecnica no invasiva, y relativamente barata, pero la cantidad de falsos positivos y falsos negativos producida por estos metodos esta muy por encima de los maximos aceptados en la practica medica. Los avances recientes en las tecnicas de imagenologia parecerian hacer posible la reduccion de estas tasas. Como consecuencia de esto, estamos asistiendo a un nuevo interes por la colonoscopia virtual. En este trabajo se presenta un sistema completo de diagnostico asistido por computadora. La primera etapa del sistema es la segmentacion, que consiste en la reconstruccion 3D de la superficie del colon a partir del volumen tomografico. El aporte principal en este paso es el suavizado de la imagen. A partir de la superficie, se detectan aquellas zonas candidatas de ser polipos mediante una estrategia multi-escala que permite delinear con precision la zona. Luego para cada candidato se extraen caracteristicas geometricas y de textura, que son calculadas tambien en el tejido que rodea la zona a efectos de compararlas. Finalmente las zonas candidatas se clasifican utilizando SVM. Los resultados obtenidos son prometedores, permitiendo detectar un 100% de los polipos mayores

Colorectal cancer is the second leading cause of cancer-related death in the United States, and the third cause worldwide. The early detection of polyps is fundamental, allowing to reduce mortality rates up to 90%. Nowadays, optical colonoscopy is the most used detection method due in part to its relative high performance. Virtual Colonoscopy is a promising alternative technique that emerged in the 90's. It uses volumetric Computed Tomographic data of the cleansed and air-distended colon, and the examination is made by a specialist from the images in a computer. Therefore, this technique is less invasive and less expensive than optical colonoscopy, but up to now the false positive and false negative rates are above the accepted medical limits. Recent advances in imaging techniques have the potential to reduce these rates; consequently, we are currently re-experiencing an increasing interest in Virtual Colonoscopy. In this work we propose a complete pipeline for a Computer-Aided Detection algorithm. The system starts with a novel and simple segmentation step. We then introduce geometrical and textural features that take into account not only the candidate polyp region, but the surrounding area at multiple scales as well. This way, our proposed CAD algorithm is able to accurately detect candidate polyps by measuring local variations of these features. Candidate patches are then classi ed using SVM. The whole algorithm is completely automatic and produces state-of-the-art results, achieving 100% sensitivity for polyps greater than 6mm in size with less than one false positive per case, and 100% sensitivity for polyps greater than 3mm in size with 2:2 false positives per case.