Automated segmentation of liver and liver cysts from bounded abdominal MR images in patients with autosomal dominant polycystic kidney disease

Abstract

Liver and liver cyst volume measurements are important quantitative imaging biomarkers for assessment of disease progression in autosomal dominant polycystic kidney disease (ADPKD) and polycystic liver disease (PLD). To date, no study has presented automated segmentation and volumetric computation of liver and liver cysts in these populations. In this paper, we proposed an automated segmentation framework for liver and liver cysts from bounded abdominal MR images in patients with ADPKD. To model the shape and variations in ADPKD livers, the spatial prior probability map (SPPM) of liver location and the tissue prior probability maps (TPPMs) of liver parenchymal tissue intensity and cyst morphology were generated. Formulated within a three-dimensional level set framework, the TPPMs successfully captured liver parenchymal tissues and cysts, while the SPPM globally constrained the initial surfaces of the liver into the desired boundary. Liver cysts were extracted by combined operations of the TPPMs, thresholding, and false positive reduction based on spatial prior knowledge of kidney cysts and distance map. With cross-validation for the liver segmentation, the agreement between the radiology expert and the proposed method was 84% for shape congruence and 91% for volume measurement assessed by the intra-class correlation coefficient (ICC). For the liver cyst segmentation, the agreement between the reference method and the proposed method was ICC = 0.91 for cyst volumes and ICC = 0.94 for % cyst-to-liver volume.

Figure 1

MRI images of polycystic livers with a range of cystic burdens.

Figure 2

The overview of the automated segmentation process. The segmentation was operated in five steps: preprocessing, spatial prior probability map construction, tissue prior probability map construction, liver segmentation, and cyst extraction.

Figure 3

Images of spatial prior probability map (SPPM) presenting (a) coronal mid cross-section and (b) 3D surface rendering with the SPPM value p_SPPM = 0.5.

Figure 4

Typical histogram of (a) T₂-weighted abdominal MRI and (b) regions corresponding to SPPM probability p_SPPM > 0.5 (solid line) with fitted probability density function (dashed line) for the liver parenchyma.

Figure 5

Tissue prior probability maps: (a) original image, (b) intensity prior probability map for liver parenchyma, and (c) shape prior probability map for cyst tissue.

Figure 6

Illustration of level set evolution in columns at (a) initial, (b) 25, (c) 50, and (d) final iterations. The zero-level sets are delineated in solid lines. The four images on the top row represent the three-dimensional rendering of the entire zero level sets over the original image. The subsequent rows from top to bottom correspond to 25%, 50%, and 75% slice location from anterior to posterior within the volume, respectively.

Figure 7

The heat map depicting the cyst extraction accuracy in terms of all feasible combination of the parameters, α and β, in which the color bar indicates the ICCs on the right.

Figure 8

Cyst extraction process: (a) initial extraction of liver cysts including a falsely segmented kidney cyst (yellow arrow), (b) distance map to remove falsely segmented kidney cysts, and (c) final extraction of liver cysts. The small circle in (b) represents the centroid of the liver mask.

Figure 9

The automated segmentation boundary (red contour) of the liver superimposed with the manual reference segmentation (green contour) in four subjects representing (a) the anterior 25% cross-section; (b) the middle section; (c) the posterior 25% cross-section of the liver; and (d) the 3D surface renderings of the segmented livers.

Figure 10

Scatter plots of liver volume measurements between the reference and automated methods at two test sets: (a) D₁ and (b) D₂. The ICCs were (a) 0.91 (P< 0.001; CI: 0.86–0.94) and (b) 0.90 (P< 0.001; CI: 0.85–0.94). The diagonal line in the figure represents the line of identity.

Figure 11

Extraction of liver cysts in four subjects from Figure 9. The extracted cysts are delineated in yellow contours. While most of clearly-defined cysts were extracted, some small faintly-defined cysts were not segmented.

Figure 12

Scatter plots of (a) cyst volume measurements and (b) volumetric percentage of cysts relative to liver between the reference and automated methods. The intra-class correlation coefficients were (a) 0.91 (P< 0.001; CI: 0.88–0.94) and (b) 0.94 (P< 0.001; CI: 0.92–0.96). The diagonal line in the figure represents the line of identity.