Medical Segmentation Using Sobolev Optical Flow

Abstract

In computer aided detection (CAD) of the pulmonary nodules, automated analysis of nodules within the complex background of anatomic structures is extremely challenging for clinicians. The identification of the lung structures is the initial stage in CAD for improving the detection sensitivity. This paper presents a novel automated lung segmentation method for nodule detection from CT images, using the information provided about motion of the tissue within the lung and pulmonary boundaries. A deformable image registration technique, optical flow, is used to detect the structures in magnitude to difference between two adjacent slices from a CT scan. Recent research has shown that L2 -type inner product introduces a pathological Riemannian metric on the space of smooth curves. Consequently, we refine our optical flow constraint in Sobolev metrics, which induce favorable regularity properties in gradient flows. Tests with real medical images demonstrate the method and its implementation.