Random forest classification and local region-based, level-set segmentation for quantitative ultrasound of human lymph nodes

Abstract

To detect metastatic foci in excised human lymph nodes (LNs) using three-dimensional (3D), high-frequency quantitative ultrasound (QUS), the 3D envelope data must be accurately segmented into LN parenchyma (LNP), fat and normal saline (NS). However, automatic segmentation of the 3D data is challenging because of speckle as well as low contrast and intensity inhomogeneities caused by focusing and attenuation effects. We describe a novel method to automatically segment these media using initialization and refinement steps. In the first step, random forest classification (RFC) is employed for initial segmentation of the 3 media. To train the forest classifier and classify each voxel, features including backscattered energy, statistical parameters and contextual information are extracted from LN envelope data. In the second step, the initialization is refined by a 3-phase, local region-based, level-set segmentation method that uses the gamma probability density function as a statistical model of the envelope data. To handle depth-dependent data inhomogeneity efficiently, the gamma distribution parameters are estimated locally in transverse slices. From a database of 54 representative LNs acquired from colorectal-cancer patients, 12 LNs were used to train the random forest classifier, and the 42 remaining LNs were used for evaluation. The Dice similarity coefficient (DSC) was used to compare automatic and manual segmentation. For RFC alone, DSCs were 0.922 ± 0.022, 0.801 ± 0.075 and 0.959 ± 0.013 for LNP, fat and NS, respectively. For initialization and refinement, significantly better DSCs were obtained: 0.937 ± 0.021, 0.824 ± 0.074 and 0.961 ± 0.009 (Wilcoxon signed rank test). Results also demonstrate that accurate QUS estimates can be obtained with automatic segmentation in excised colorectal LNs, thus eliminating the need for operator-dependent, manual segmentation.