Automated CT Based Liver Volume Assessment - Definiens

Automated CT Based
Liver Volume Assessment
Bilal A. Ahmed 1 , Rene Korn 2 , Markus Kietzmann 2 ,
Johann Kim 2 , Guenter Schmidt 2 , Gerd Binnig 2 ,
Masoom A. Haider 1
1 Joint Department of Medical Imaging, University Health Netowork,
Mount Sinai Hospital and Womens College Hospital.
University of Toronto, Toronto, Canada
2 Definiens AG, Munich, Germany
RSNA 2009

Background
• Liver volume is an important measure of functional
capacity
• Much prior work has been done in normal livers without
metastases
• It is routinely used for surgical planning for liver
resections
• Manual volumetry is time-consuming, tedious and
generally not reproducible
• Therefore, there has been a growing interest in the
development of fast and accurate segmentation methods
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Purpose
To evaluate the accuracy of liver volume
measured by a fully-automated
segmentation software that quantifies liver
volume in patients with metastatic liver
disease using contextual information from
adjacent structures, including lungs, the
spine, ribs and gallbladder
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Methods
• 30 consecutive CT scans from distinct patients with liver
metastases were obtained. Patients had tumor burdens form
minimal to heavily occupying much of the liver
– 320 slice Toshiba CT scanner
– Study details:
• 1mm collimation
• Portal venous phase contrast-enhanced scans of the abdomen
• All scans were manually segmented to define reference standard
for liver boundaries (including metastatic lesions) using in-
house software
– Parts of the inferior vena cava and portal vein were included within
the liver boundary when enclosed by liver parenchyma in order to
make the liver contour smooth
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Methods
• Five representative cases were selected as a
training set for the automatic segmentation
algorithm
• The remaining 25 cases were used to test the
refined algorithm
• Resultant automatic segmentations generated were
compared to the manually generated reference
segmentations and deviations were measured
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Automatic Segmentation Algorithm
• Fully automatic segmentation algorithm developed in the Cognition Network
Language (CNL) with Definiens Developer XD 1.0
• Automatic pre-processing using rule-based segmentation
– Defines a set of rules which is used to successively extract different
structures from the test images
– Order of segmentation/classification is: background and body � lungs
and other intra-body air � fat, muscle and bones � spine and ribs
– Incorporates knowledge about intensity distributions, neighborhood
relations and object geometry
• Automatic selection of a seed based on geometry, position and size
relationship
– Order of segmentation/classification is: liver seed � gall bladder
• Automatic seed growing
– Growing process is halted by previously detected contextual structures
– Order of segmentation/classification is: liver seed � liver
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Automatic Segmentation Algorithm
Automatic Preprocessing
Automatic Liver Seed Selection
Algorithm: Liver Refinement/Growing
Original Image updating
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Automatic Segmentation Algorithm
Preprocessing
1. Background is separated from the body
2. Right and left lungs are identified as context objects
3. Based on size and intensity (HU), additional context objects are
generated (e.g. fat, muscle, other organs, bones)
4. Within the bone objects, spine and ribs are identified as additional
context objects based on ROI (using lungs) and shape
5. An additional layer of 3D edge information is calculated
Note: Context objects do not have to be 100% perfectly segmented to guide the analysis.
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Based on
intensity and
size the
background is
separated from
the body
Algorithm: Preprocessing
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Algorithm: Preprocessing
Based on
intensity,
position and
size right and
left lungs are
identified as
context objects
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Algorithm: Preprocessing
Based on size
and intensity
(HU),
additional
context objects
are generated
(e.g. fat,
muscle and
bones)
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Algorithm: Preprocessing
Within the
bone objects,
spine and ribs
are identified
as additional
context objects
based on ROI
(using lungs)
and shape
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Algorithm: Liver Seed
Liver Seed
1. Exploiting the 3D edge layer, potential liver seeds within the
muscle context objects are identified
2. Based on size, a single liver seed is automatically selected
3. Based on right lung and liver seed, the gallbladder context
object is segmented
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Potential
liver seeds
are identified
within the
muscle
context
objects
Algorithm: Liver Seed
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Based on
size, a single
liver seed is
automatically
selected
Algorithm: Liver Seed
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Based on
muscle, right
lung, liver
seed and
size, the
gallbladder
context
object is
identified
Algorithm: Liver Seed
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Algorithm: Liver refinement/Growing
Liver Refinement/Growing
1. The liver seed is split into an upper and lower part
2. First, the lower liver seed is grown (gallbladder, ribs,
edge)
3. Second, the upper liver is grown (ribs, lung wings, edge)
Update Original Image
Results are transferred back to original image size and refined
(prior to this, all work was done based on a downscaled image by a factor of 0.5)
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Algorithm: Liver Refinement/Growing
The liver
seed is split
into an upper
and lower
part
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Algorithm: Liver Refinement/Growing
First, the
lower liver
seed is
grown
(gallbladder,
ribs, edge)
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Algorithm: Liver Refinement/Growing
Second, the
upper liver
seed is
grown (ribs,
lung wings,
edge)
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Algorithm: Update Original Image
Results are
transferred to
original
image size
(100%) and
refined
further
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Results: Measurements
• Two measurements were used to evaluate segmentation accuracy:
– Minimum overlap
• A, B Sets of voxels from tracing and classification data respectively
– Correlation Coefficient
• tp = number of true positive voxels
• fp = number of false positive voxels
• tn = number of true negative voxels
• fn = number of false negative voxels
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Results: Sample Segmentation I
Statistics
• CC 0.86
• MO 86%
• 9:47 min
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Results: Sample Segmentation II
Statistics
• CC 0.94
• MO 89%
• 6:29 min
(best case)
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Results: Summary
• Correlation coefficient between the measured and
estimated volumes: 0.864 ± 0.04 (0.777 – 0.935)
• Mean volumetric overlap: 77% ± 6 (64 – 88%)
• Mean liver volume measured by the automatic
segmentation algorithm: 1820 cm³ ± 788 (710 – 4565 cm³)
• Mean liver volume measured by manual segmentation:
1806 cm³ ± 685 (934 – 3460 cm³)
• Mean total time for the automatic segmentation: 4.8
minutes ± 1.2 on E6750 @2.66GHz Intel Core 2 Duo
CPU with 3.25 GB RAM
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Discussion
• Initial evaluation of fully automated algorithm using
contextual information from surrounding structures
shows promising results
• Small training set (5 cases) and large test set (25 cases)
is challenging test of performance
• Unique advantage of this approach is a universal toolkit
that can be used to develop highly specific algorithms
detecting all kinds of objects
– Has wide array of applications in image segmentation
• Further evaluation will include specific segmentation of
liver metastases
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

Conclusion
• Fully-automated segmentation software that
quantifies liver volumes using contextual
information from adjacent structures
exhibits promising accuracies in acceptable
time frames in patients with varying degrees
of metastatic disease
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto

References
• Heimann T, van Ginneken B, Styner MA, et al. Comparison and evaluation of methods
for liver segmentation from CT datasets. IEEE Trans Med Imaging. 2009
Aug;28(8):1251-65.
• Schmidt G, Athelogou MA, Schonmeyer R, Korn R, and Binnig G. Cognition network
technology for a fully automated 3D segmentation of liver. Proc MICCAI Workshop on
3D Segmentation in the Clinic: a Grand Challenge. 2007; pp. 125–133.
• Prasad SR, Jhaveri KS, Saini S, Hahn PF, Halpern EF, Sumner JE. CT tumor
measurement for therapeutic response assessment: comparison of unidimensional,
bidimensional, and volumetric techniques initial observations. Radiology 2002;
225:416-419.
• Lamecker H, Lange T, Seebass M, Eulenstein S, Westerhoff M, Hege HC. Automatic
segmentation of the liver for preoperative planning of resections. Stud Health Technol
Inform 2003; 94:171-173.
• Selver MA, Kocaoglu A, Demir GK, Dogan H, Dicle O, Guzelis C. Patient oriented and
robust automatic liver segmentation for pre-evaluation of liver transplantation. Comput
Biol Med 2008; 38:765-784.
• Athelogou, M. et al.: Definiens Cognition Network Technology – a novel multimodal
image analysis technique for automatic identification and quantification of biological
image contents. In Shorte, S.L., Frischknecht, F., eds.: Imaging Cellular and Molecular
Biological Functions. Springer-Verlag, Berlin Heidelberg (2007) 407-421.
Medical Imaging – Princess Margaret Hospital – University Health Network – Mount Sinai Hospital – University of Toronto