3D kidney segmentation from abdominal diffusion MRI using an appearance-guided deformable boundary

Abstract

A new technique for more accurate automatic segmentation of the kidney from its surrounding abdominal structures in diffusion-weighted magnetic resonance imaging (DW-MRI) is presented. This approach combines a new 3D probabilistic shape model of the kidney with a first-order appearance model and fourth-order spatial model of the diffusion-weighted signal intensity to guide the evolution of a 3D geometric deformable model. The probabilistic shape model was built from labeled training datasets to produce a spatially variant, independent random field of region labels. A Markov-Gibbs random field spatial model with up to fourth-order interactions was adequate to capture the inhomogeneity of renal tissues in the DW-MRI signal. A new analytical approach estimated the Gibbs potentials directly from the DW-MRI data to be segmented, in order that the segmentation procedure would be fully automatic. Finally, to better distinguish the kidney object from the surrounding tissues, marginal gray level distributions inside and outside of the deformable boundary were modeled with adaptive linear combinations of discrete Gaussians (first-order appearance model). The approach was tested on a cohort of 64 DW-MRI datasets with b-values ranging from 50 to 1000 s/mm2. The performance of the presented approach was evaluated using leave-one-subject-out cross validation and compared against three other well-known segmentation methods applied to the same DW-MRI data using the following evaluation metrics: 1) the Dice similarity coefficient (DSC); 2) the 95-percentile modified Hausdorff distance (MHD); and 3) the percentage kidney volume difference (PKVD). High performance of the new approach was confirmed by the high DSC (0.95±0.01), low MHD (3.9±0.76) mm, and low PKVD (9.5±2.2)% relative to manual segmentation by an MR expert (a board certified radiologist).

Fig 1. The proposed framework for kidney segmentation.

Fig 2

Typical coronal cross-section DW-MRI samples showing (a) low contrast between the kidney and surrounding abdominal tissues at b₀; (b) inter-patient anatomical differences at b₀, (c) low signal-to-noise ratio (SNR), especially, at higher b-values (e.g., b₁₀₀₀); (d) image artifacts; and (e) geometric distortion/diffused boundaries.

Fig 3. 3D zero-level set of a function Φ(p = [x, y, z], t).

Fig 4. The nearest 26-neighborhood of a voxel for the 4^th-order spatial model and examples of its 2^nd-order cliques (upper raw), 3^rd-order cliques (middle raw), and 4^th-order cliques (lower raw).

Note that the central voxel is shown in yellow, while its neighbors are shown (i) in red for the same plane and (ii) in blue and purple for the adjacent planes.

Fig 5. 3D co-alignment of training DW-MRI datasets (S₁:S_N) to a single reference: The first and second rows present the overlapped 3D kidney volumes before and after the alignment, respectively.

Note that the reference subject appears in magenta, while the targets are shown cyan.

Fig 6. Our segmentation (red) with respect to the expert’s manual ground truth (green): The coronal (left), axial (middle), and sagittal (right) cross-sections for two different subjects in the first and second rows.

Fig 7. Our segmentation (red) with respect to the expert’s manual ground truth (green) using the 4^th-order MGRF (first row) compared to our previous segmentation using the 2^nd-order MGRF (second row) [54], [55] for three different subjects (columns), where the first, second, and third columns show large, moderate, and small differences in yellow regions (false positive (FP)) and blue regions (false negative (FN)), respectively.

Fig 8. Comparative cross-sectional segmentation results for our approach (a), the vector level sets [35] (b), the segmentation using the Random Forest classifier [42] (c), and the traditional CV [31] level set (d) for two independent subjects (rows).

The model segmentation is shown in red with respect to the manual ground truth (green) from an expert.

Fig 9. Our 3D segmentation (red) with respect to the expert’s manual ground truth (green) for three subjects with the associated accuracy scores.

Fig 10. Comparative coronal cross-sectional segmentation results for the proposed approach (a), the vector level sets [35] (b), the segmentation using the Random Forest classifier [42] (c), and the traditional CV [31] level set (d) for one subjects at b₀ (first row) and higher b_i values (b₅₀₀ (second raw) and b₁₀₀₀ (third raw)).

The model segmentation is shown in red with respect to the manual ground truth (green) from an expert.