Fig 1.
Flow chart of the image segmentation scheme.
The proposed segmentation algorithm includes five consecutive steps: DCE micro-CT images acquisition, data dimension reduction, supervoxel generation, supervoxel classification, and target organs’ extraction.
Fig 2.
Dynamic contrast enhancement procedure after contrast agent administration.
(A) Representative coronal micro-CT images before contrast agent injection and at 0 s, 50 s, 100 s, 150 s, and 200 s post-contrast injection. The image at 0 s was acquired during the inflow of contrast agent. All of the images are displayed with the same gray scale window. (B) The relative signal enhancement versus time curves of regions depicted by the arrows in (A) with the same colors.
Fig 3.
Results of supervoxel generation shown in 2-D axial slices.
The green denotes the boundaries of supervoxels. (A) Axial slice of thorax containing the heart and lung. (B) Axial slice of upper abdomen containing the liver and spleen. (C) Axial slice of lower abdomen containing the kidney and intestine.
Table 1.
Number of supervoxels for each category chosen to constitute the total data set.
Fig 4.
Influence of the number of training samples on segmentation accuracy.
10%, 30%, 50%, 70%, and 90% of the total samples of each category were selected randomly from the total data set and consisted of training sets for classification. Each case was repeated five times. The means and standard deviations of Dice similarity coefficients for the heart, liver, spleen, lung, and kidney were calculated (compared with M1). (A) The DSC of the organs classified by SVM. (B) The DSC of the organs classified by RF. For the lung, one value at 10% and two values at 30% were excluded from statistics because it failed to extract the lung by post-processing.
Fig 5.
Visual comparison of the segmentation results with the reference datasets, shown in 3-D isosurface rendering.
Left column: left lateral view. Right column: posterior view. Top row: manual segmentation (M1). Middle row: segmentation obtained by the SVM. Bottom row: segmentation obtained by the RF.
Fig 6.
Visual comparison of the segmentation results with the reference datasets, shown in 2-D images.
The organ boundaries of manual segmentation (M1) and automatic segmentation based on the SVM and the RF are superimposed on two coronal images (A, B) and two sagittal images (C, D).
Fig 7.
Quantitative evaluation of the proposed methods by comparison to manual segmentation.
‘SM1’ and ‘RM1’ represents the comparison of the automatic segmentation by SVM and RF with the manual segmentation (M1), respectively. ‘M1M3’ compares the manual segmentations of two independent experts. ‘M1M2’ compares two manual segmentation repetitions of one expert. (A) Dice similarity coefficient. (B) False positive ratio. (C) False negative ratio. (* Indicates p < 0.05.)