3D level set esophagus segmentation in thoracic CT images using spatial, appearance and shape models

Sila Kurugol, Jennifer G. Dy, Gregory C. Sharp, Dana H. Brooks

Abstract: We propose a 3D segmentation algorithm to locate the esophagus in thoracic CT scans using a learning based approach. To ease the training data requirement and allow maximum inter-subject flexibility, we built a simple algorithm based on normalization to anatomical reference points to match a training set of thoracic CTs instead of a full statistical registration based on neighboring structures. We use spatial and appearance models to locate the centerline. We build a shape model by subtracting the centerline and applying PCA to the training data sets. The shape model includes a mean shape plus the weighted combination of modes. To locate the esophageal wall, we optimize a cost function including terms for appearance, shape model, smoothness constraints and air/contrast model using a 3D level set framework.