Hierarchical Framework for Automatic Pancreas Segmentation in MRI Using Continuous Max-Flow and Min-Cuts Approach

摘要

Accurate, automatic and robust segmentation of the pancreas in medical image scans remains a challenging but important prerequisite for computer-aided diagnosis (CADx). This paper presents a tool for automatic pancreas segmentation in magnetic resonance imaging (MRI) scans. Proposed is a framework that employs a hierarchical pooling of information as follows: identify major pancreas region and apply contrast enhancement to differentiate between pancreatic and surrounding tissue; perform 3D segmentation by employing continuous max-flow and min-cuts approach, structured forest edge detection, and a training dataset of annotated pancreata; eliminate non-pancreatic contours from resultant segmentation via morphological operations on area, curvature and position between distinct contours. The proposed method is evaluated on a dataset of 20 MRI volumes, achieving a mean Dice Similarity coefficient of 75.5 ± 7.0% and a mean Jaccard Index coefficient of 61.2 ± 9.2%.