Segmentation of liver, its vessels and lesions from CT images for surgical planning
 Dário AB Oliveira^{1}Email author,
 Raul Q Feitosa^{1} and
 Mauro M Correia^{2}
https://doi.org/10.1186/1475925X1030
© Oliveira et al; licensee BioMed Central Ltd. 2011
Received: 1 December 2010
Accepted: 20 April 2011
Published: 20 April 2011
Abstract
Background
Cancer treatments are complex and involve different actions, which include many times a surgical procedure. Medical imaging provides important information for surgical planning, and it usually demands a proper segmentation, i.e., the identification of meaningful objects, such as organs and lesions. This study proposes a methodology to segment the liver, its vessels and nodules from computer tomography images for surgical planning.
Methods
The proposed methodology consists of four steps executed sequentially: segmentation of liver, segmentation of vessels and nodules, identification of hepatic and portal veins, and segmentation of Couinaud anatomical segments. Firstly, the liver is segmented by a method based on a deformable model implemented through level sets, of which parameters are adjusted by using a supervised optimization procedure. Secondly, a mixture model is used to segment nodules and vessels through a region growing process. Then, the identification of hepatic and portal veins is performed using liver anatomical knowledge and a vein tracking algorithm. Finally, the Couinaud anatomical segments are identified according to the anatomical liver model proposed by Couinaud.
Results
Experiments were conducted using data and metrics brought from the liver segmentation competition held in the Sliver07 conference. A subset of five exams was used for estimation of segmentation parameter values, while 15 exams were used for evaluation. The method attained a good performance in 17 of the 20 exams, being ranked as the 6^{th} best semiautomatic method when comparing to the methods described on the Sliver07 website (2008). It attained visual consistent results for nodules and veins segmentation, and we compiled the results, showing the best, worst, and mean results for all dataset.
Conclusions
The method for liver segmentation performed well, according to the results of the numerical evaluation implemented, and the segmentation of liver internal structures were consistent with the anatomy of the liver, as confirmed by a specialist. The analysis provided evidences that the method to segment the liver may be applied to segment other organs, especially to those whose distribution of voxel intensities is nearly Gaussian shaped.
Background
In medical image analysis, imageguided surgery and organ visualization, segmentation is a crucial step. The segmentation process is particularly arduous in abdominal computer tomography (CT) images because different organs lie within overlapping intensity ranges and are often near to each other anatomically. Therefore, usually it is not possible to define accurately the boundaries of organs, their vessels and lesions using simple threshold based segmentation. On the other hand more complex algorithms involve comparatively many parameters of which adjustment is not a simple issue.
Numerous techniques have been proposed in the literature for extraction of organ contours in abdominal CT scans. They can be roughly divided in two main groups: model driven and data driven approaches [1].
Model driven techniques (e.g. [2, 3]) use predefined models to segment the meaningful objects in the images being analyzed. In this kind of technique a model describing the organ to be segmented is defined in terms of object characteristics such as position, texture and spatial relation to other objects, and the algorithm searches the images for instances that fit the given model.
Data driven techniques (e.g. [4, 5]) try to emulate the human capacity of identifying objects using some similarity information present on image data, automatically detecting and classifying objects and features in images. Many of them use traditional techniques such as region growing and thresholds, combined with some prior knowledge about the object being analyzed.
Level set methods [6] are model driven methods that rely on partial differential equations to model deforming isosurfaces. These methods have been used successfully in medical image processing but usually require human intervention to set an initial solution and indicate explicitly when the model should stop expanding. Moreover, semi automatic level set based methods involve a time consuming trial and error procedure for optimum parameter tuning.
The parameters in the implementation of traditional level sets are related to the curves mean curvature, propagation rate and advection of the curve to certain characteristics of the image. The manual definition of these values on level set methods is a complex task, because their relation with the final result is unclear and there is no guarantee that the optimal set of values will be found. Therefore, there is a demand for methods to define such parameters automatically.
Some works approaching liver segmentation using level set based methods are found on the literature. In [7] a level set method without edges was proposed to segment the liver, using the ChanVese methodology ([8]). In [9] an active model based on level sets was proposed to segment the liver, using a multiresolution concept to reduce processing time. In both works, in spite of achieving good results, the parameters were not automatically defined, and the segmentation did not included liver vessels and lesions.
In this work we propose a complete methodology to segment the liver ([10–12]) and its internal structures, such as vessels ([11]) and nodules, using level sets, stochastic optimization, and Gaussian mixture model. It also proposes a methodology to separate the liver into segments according to the Couinaud [13] anatomical model.
The subsequent text is organized in the following way. First, the liver anatomy is briefly described. The theoretical fundamentals of level sets and the optimization algorithm used are then presented. Next, we present the proposed liver segmentation method in details and the parameter estimation procedure. Then, the segmentation of vessels and nodules, the identification of portal and hepatic veins and the segmentation of Couinaud regions are described. Last, we report the results of the corresponding experimental performance evaluation, and present the main conclusions.
Liver Anatomy
The liver is a vital and complex gland, with a wide range of functions, including detoxification, protein synthesis, and production of biochemicals necessary for digestion. It has two distinct descriptions, according to its morphological and functional aspects. Morphologically the liver presents 4 lobes: left, right, caudate and quadrate. Functionally the liver present a different anatomy, and one of the proposed ones is the Couinaud model, which is used within this work.
The functional anatomy model of Couinaud proposes the division of the liver into eight different regions according to the portal and hepatic veins positions:
Segment I  caudate/Spiegel lobe
Segment II  left posterolateral
Segment III  left anterolateral
Segment IVa  left superomedial
Segment IVb  left inferomedial
Segment V  right anteroinferior
Segment VI  right posteroinferior
Segment VII  right posterosuperior
Segment VIII  right anterosuperior
Level Sets
The level set methods theory was originally developed by Sethian and Osher [6, 14] for analyzing and computing moving fronts in a variety of different settings. These methods were firstly introduced in medical imaging by Malladi et al. [15].
The level set methodology relies on a numerical technique that uses level set function to model and compute continuous deformable models [16] using an implicit representation. It has been applied to different areas, such as computer graphics, computer vision and image processing. In image processing it was successfully applied to segment structures with boundaries not clearly defined, that usually cannot be segmented using simple data driven methods, based on thresholds and line extraction. The level set approach takes into account other properties such as mean curvature to model the target object, in such a way that it is possible to avoid undesirable leaks toward neighbouring structures, preserving the boundaries of the segmented object even when they are not clearly defined through the entire object.
The main idea of level sets is to embed a deformable model in a d+1 dimensional space, in order to segment iteratively an object in a d dimensional space. The main advantage of level sets is that it allows changes of surface topology implicitly. As it embeds the evolving surface, also called interface, in a higher dimensional function, the interface can split into several connected components or merge from different connected components naturally. It occurs because the embedding level set function, which is used to deform the surface, remains continuous.
Considering ψ(x,t) the level set function, x the position vector and t the time step of the level set evolution, the evolving surface is represented as the zero level set of ψ(x,t) = 0. This time step controls the evolution of the interface in time. The interface ψ(x,t+1) = 0 is computed taking into account the interface of the previous time step ψ(x,t) = 0, using partial differential equations. The segmentation procedure ends when the RMS difference between ψ(x,t) = 0 and ψ(x,t1) = 0 is less than a predefined minimum RMS value.
The level set function is usually a smooth well behaved function, for numerical stability reasons during partial differential equations computation. This work uses the signed distance function that computes the distance of each voxel to its closest voxel in the interface. This distance is negative inside the interface, and positive outside, in such a way that the position of a given point in relation to the interface can be easily determined by looking at the sign of the level set function in that point.
Note that the gradient Δψ (or its module) of ψ(x,t) appears in all terms of the equation. As ψ(x,t) was defined as the signed distance function, the gradient of ψ(x,t) points from inside to outside considering the interface ψ(x,t) = 0, and Δψ = 1, by definition. It defines, together with other parameters, the direction and magnitude of interface evolution.
The functions A, P and Z are usually calculated from the input image. A is a vector field responsible for the advection term. It attracts the evolving surface to determinate features, usually related to boundaries of objects, and predefined barriers. It is weighted by the constant advection weight α, and multiplied by the gradient of ψ(x,t).
P is a propagation image, also called speed image, which controls how the level sets propagates through the coordinate space. This image normally has high values in regions where the interface can expand quickly, and values close to zero in regions where it should move slowly or stop (normally close to important features). It is weighted by the constant propagation weight β and multiplied by the module of the gradient of ψ(x,t).
К is the mean curvature of the interface, and is defined as the divergence of the normal to the interface. It is usually calculated using first and second derivatives of the interface, based on finite differences. In this way, К > 0 for convex regions, and К < 0 for concave regions. Z is a spatial modifier for the mean curvature К, and modifies the value of К in a determinate spatial position. In this work P was defined in such a way that the curvature has less importance close to important features. It is weighted by the constant mean curvature weight γ and multiplied by the module of the gradient of ψ(x,t).
The basic level set model has these three terms but a segmentation algorithm based on level sets may add new different terms, or omit one or more terms, depending on the problem complexity.
Supervised adaptation of segmentation parameters
Many of the parameters mentioned until now must be properly adjusted for the level set method to produce accurate results. The determination of appropriate parameter values usually cannot be done using heuristics mainly due to the complexity of the target application. It is generally also not possible to determine the optimum values using nonlinear optimization methods for parameter adjustment, due to multiple local minima.
A number of automatic methods have been proposed to optimize parameters in such cases. They can be grouped into two categories: supervised and unsupervised [17]. Unsupervised quality assessment methods, e.g., [18] and [19], rely on statistics computed from the image being segmented and the segmentation itself. Supervised methods try to avoid subjectivity through the use of reference segments, which represent an ideal segmentation. Quality is assessed by measuring the deviation of the segmentation outcome from the references.
Stochastic approaches for optimization are options to consider when calculusbased optimization methods cannot be applied, such as in our methodology. Although stochastic methods are computationally intensive and do not guarantee that the optimum will be reached, they are able to work with virtually any objective function.
In the present work we used a genetic algorithm (GA), which is a stochastic technique inspired by the evolution theory and widely applied in many optimization problems [20].
GA methods aim at automatically adjusting segmentation parameters, taking into consideration a set of reference segments, delineated by an interpreter. The basic idea is to search the parameter space for a set of values that optimizes a given fitness function, which expresses numerically the discrepancy among the automatically generated objects and the references. For details refer to [21, 22].
It is worth emphasizing that the model described hitherto is not bound to any particular parameter optimization technique. There may be optimization approaches other than GA (e.g. simulated annealing) that could be more efficient for this particular problem. An evaluation of possible optimization techniques that could be used to get the best parameters goes beyond the scope of this paper, but is an interesting issue to be pursued in the continuation of this research. Nevertheless, the results reported attest that our method may achieve good results using a GA for parameter estimation and some approaches for different image segmentation methods were already published (e.g. [23]) with encouraging results.
Methods
The methodology consists of four steps executed sequentially. Firstly the liver is segmented using level sets with optimal parameters computed by a GA. The level set segmentation uses an initial userdefined liver segment in one slice, and then segments the liver through all other slices, using a Gaussian fit to define the speed image where the level sets propagates. The initial solution at each slice is defined as the region previously segmented on an adjacent slice.
Secondly, the vessels and nodules are segmented using a Gaussian mixture model. In this work only the nodules that appear darker than liver parenchyma, the adipose ones, are contemplated. Thus, the model consists in discriminating Gaussians in a mixture model considering that the vessels appear brighter than the liver parenchyma and the nodules appear darker. Then a region growing method using the information of each Gaussian implements the segmentation.
Thirdly, the vessels are classified into portal vein or hepatic vein using a simple vein tracking method, and finally, a geometrical approach is applied to split the liver into different regions of Couinaud, using the anatomy of these two defined veins.
These steps are described in further details in the following sections.
Liver Segmentation
The proposed method relies on two hypotheses: the liver parenchyma is roughly homogeneous, and liver veins are mainly inside the liver, as well as liver nodules. The impact of these heuristics on cases where peripheral nodules and veins are present is discussed in details in the section on experimental results.
Firstly an initial solution is defined manually in just one slice. This initial solution is expected to contain great part of the liver, but does not need to be accurate, as it will be later deformed too. Then an iterative process takes place, both upwards and downwards through the whole exam, in which the liver is segmented at each slice sequentially. In this process the initial solution of a new slice to be processed is the previously segmented result of the adjacent slice, which is then deformed towards the liver boundaries using the approach based on level sets, with some of the parameters being adjusted at each slice.
In this work, the advection term was not used, because it is closely related to boundary features and the boundaries of the liver cannot be robustly distinguished from its anatomical neighbouring structures.
We assume that the liver parenchyma is roughly homogeneous and it is possible to define two thresholds T _{ L }and T _{ H }that define a gray level range where most of parenchyma voxels lie on.
The propagation P(x) in the preceding equation is positive when the pixel intensity lies inside the range [T _{ L } ,T _{ H }] and negative when it is outside the range. It implies that the surface expands in regions where pixel values are inside the range, and shrinks otherwise. It can be further observed in equation 2 that the surface evolving speed decreases or even becomes zero, in regions where pixel values are close to T _{ L }and T _{ H }. In this sense, considering that T _{ L }and T _{ H }are values close to the ones observed on the liver boundaries, the surface evolution stops when it achieves the liver boundaries.
Clearly, the choice of proper values for T _{ L }and T _{ H }is crucial for good accuracy. We propose a simple algorithm to set up T _{ L }and T _{ H }, relying on the assumption that the voxel intensities inside the liver follow a Gaussian shaped distribution.
The mean curvature term is defined by the spatial modifier Z, the mean curvature К and the weight γ. The spatial modifier Z was defined like P, in such a way that the importance of the mean curvature К decreases close to the boundaries. This avoids that boundary details are missed by smoothing too much the surface on boundaries of the segmented object.
Once the values for G _{ L }, G _{ H }, β and γ are selected, the level set process can start at a given slice, deforming the initial solution towards the liver boundaries until the convergence criterion is achieved (minimum RMS).
The level set segmentation procedure confine the voxels inside the liver parenchyma, since their intensities meet the estimated Gaussian distribution, but nodules and veins, which normally appear respectively as darker and brighter regions are not captured. This problem is partially eliminated by the use of a 'fillholes' morphologic algorithm [26], that merges to the result of the preceding step veins and nodules that are totally inside the liver. However the problem persists when a nodule or vein is at the periphery, when they usually do not appear in the final result.
This algorithm is applied sequentially for each slice in the exam, and ends when it achieves the first and last slices, or when an initial segment is vanished by the level set algorithm in a given slice, indicating a limit of the liver.
Parameter Estimation
Processing Scheme
In the devised GA each individual consists of a set of segmentation parameter values; each parameter is represented by a gene. The fitness of each solution (individual) is calculated by comparing the segmentation produced by the solution with the reference segmentation, using five different measures of performance [27] that will be further presented.
 1.
Minimum RMS: the convergence criterion of level set function, defined in terms of the root mean squared (RMS) change in the level set function.
 2.
G _{ L }: Gaussian low factor
 3.
G _{ H }: Gaussian high factor
 4.
β: level set propagation weight
 5.
γ: level set mean curvature weight
Initially a set of solutions (initial population) is generated by computing each parameter value (genes) randomly, within given ranges. As the evolutionary process goes on, the best solutions (fittest individuals) are selected and new solutions (generations) are created from them (reproduction).
The fittest individuals have a larger probability of being selected for reproduction. Furthermore, the best individuals from one generation are kept in the next generation. The evolutionary process stops after a fixed number of generations, and the gene values of the fittest individual are taken as the final (adapted) segmentation parameter values.
Fitness Evaluation
The fitness function devised in this work is defined as the mean of five scores derived from five disparity metrics proposed by SLiver07 conference [27] that express the level of agreement between the result and the reference surfaces. These metrics try to numerically compare two different segmentations, represented by two different surfaces. In this way, they are measures of spatial differences between the two surfaces. The scores computed over these metrics take into account the uncertainty of a human specialist, in such a way that the score gives the level of agreement between a segmentation outcome and the one produced by a human specialist.
 1.
"Volumetric overlap (VOE): Is the number of voxels in the intersection of segmentation and reference, divided by the number of voxels in the union of segmentation and reference."
 2.
"Relative absolute volume difference, in percent (RVD): 1 minus the total volume of the segmentation divided by the total volume of the reference"
 3.
"Average symmetric absolute surface distance, in millimetres (ASD): the border voxels of segmentation and reference are determined and for each voxel in these sets, the closest voxel in the other set is determined (using Euclidean distance). All these distances are stored, for border voxels from both reference and segmentation. The average of all these distances gives the averages symmetric absolute surface distance."
 4.
"Symmetric RMS surface distance, in millimetres (RMSSD): is similar to the previous measure, but stores the squared distances between the two sets of border voxels. After averaging the squared values, the root is extracted and gives the symmetric RMS surface distance."
 5.
"Maximum symmetric absolute surface distance, in millimetres (MSD): is similar to the previous two, but only the maximum of all voxel distances is taken instead of the average."
The scores are calculated based on a number of sample segments manually drawn by specialists for each exam. The reference for each sample is obtained by computing the mean surface among the segmentation produced by all specialists (for details refer to [27]). For each metric the mean of the absolute differences between each specialist result to that reference is computed. These values are taken as the mean human errors. An algorithm with differences exactly equal to these values receives score of 75 points. Segmentation exactly equal to the reference receives the score of 100, and the score decreases linearly as the error increases.
Nodule and Vessel Segmentation
To segment nodules and vessels we propose a region growing method based on a Gaussian mixture model. Our approach considers three distinct object classes inside the liver: adipose nodules, liver parenchyma and vessels, which may be differentiated by their gray levels. In this work only the adipose nodules, which are less dense than liver parenchyma, are considered.
Mixture Model
where l is an index for left Gaussian, c for centre Gaussian and r for right Gaussian; and K _{ i }is a constant that multiplies the Gaussian function, with the parameters μ _{ i }and σ _{ i }.
 1.
Estimating G _{ c }
G _{ c }is considered dominant in the mixture model, i.e., K _{ c }>> {K _{ r } ,K _{ l }}. Under this assumption it is possible to identify G _{ c }using the following procedure.
 2.
Estimating G _{ l }
 3.
Estimating G _{ r }
 4.
Final adjustment
The LevenbergMarquadt optimizer is applied to find the parameter values that deliver the best fitting of G _{ l } +G _{ c } +G _{ r }to H _{ i }.
Segmentation of Nodules and Vessels using the mixture model
To segment nodules and vessels a region growing strategy is proposed that bases on the mixture model. The nodules are segmented using the estimated functions G _{ l }and G _{ c }respectively for the nodules and for the liver parenchyma. Spatial information is used to separate nodules from parenchyma when the intensity is in the range where G _{ l }and G _{ c }overlap. The region growing approach uses two different thresholds (T _{ L } ^{ lc }and T _{ H } ^{ lc }) in such a way that voxels with intensity lower than T _{ L } ^{ lc }are assigned to nodules, and are used as seeds of a process that aggregates any neighbouring voxel with intensity lower than T _{ H } ^{ lc }.
The same approach is used to segment vessels. The vessels inside the liver are brighter than liver parenchyma and have intensity distributed according to G _{ r }. Again two different thresholds are used to perform a region growing segmentation: T _{ L } ^{ rc }and T _{ H } ^{ rc }. The voxels with intensity values greater than T _{ H } ^{ rc }are classified as vessels straightforwardly and used as seeds for the region growing process that aggregates the neighbouring voxels with intensity value greater than T _{ L } ^{ rc }.
T _{ H } ^{ rc }is set empirically as the intensity value in which the proportion P^{ rc }of voxels belonging to G _{ r }and voxels belonging to the mixture G _{ r } + G _{ c }is at least of 70%. T _{ L } ^{ rc }is set as the intensity value where G _{ r }and G _{ c }intercept (T _{ H } ^{ rc }> T _{ L } ^{ rc }). Figure 6 also illustrates the procedure.
Identification of Portal and Hepatic veins
Once the vessels are segmented it is necessary to assign them to the hepatic and portal veins, which are the basis for the Couinaud segmentation.
It is assumed that three main branches compose the hepatic veins (which occur in 80% of people). Besides, it is assumed that these branches have predominantly vertical direction, coming from the upper part of the liver to the bottom part.
The hepatic main branches are identified without considering their bifurcations. A vessel is classified as one of the main branches of the hepatic vein, when it is longer than 15% of the liver height. This value was established after testing several different values.
 1.
The first slice k of I _{ vessels }containing vessels is selected.
 2.
The biggest connected component from the slice k is selected and labelled as candidate for hepatic vein main branch.
 3.
The connected component of the posterior slice with the biggest overlap with the main branch candidate is selected and merged with it.
 4.
Step 3 is repeated until no more overlapping is observed.
 5.
If the main branch candidate height is bigger than 15% of the liver height, then this candidate is accepted, otherwise it is discarded, and the process restarts.
 6.
Steps 1 to 5 are repeated until three main branches have been found.
The portal vein is identified as the biggest connected component among the vessels when the hepatic veins are extracted. The remaining vessels are merged to the hepatic branches they are connected to, or classified as auxiliary vessels, when they don't.
Couinaud Liver Segmentation
The division of the liver into Couinaud segments is based on liver anatomy. The geometric approach proposed in this work creates four planes: three approximately vertical, following the orientation of each hepatic vein main branch and one approximately horizontal, following the orientation of portal vein.
To determine each of these planes a simple algorithm of geometric fit, based on least squares method, is proposed. The voxels belonging to each branch of hepatic vein and to the portal vein are selected and a specific plane with the characteristics of the model (near vertical or horizontal) is computed. The four computed planes determine the limit of the 8 Couinaud regions.
Results and Discussion
Liver Segmentation
In order to evaluate the performance of the proposed method, a software prototype was developed in C++/C#. The prototype uses the ITK library [28] which implements the level set framework used in the segmentation method. It is also possible to visualize the results in different slice directions, as well as the 3D organ models. An additional movie showing the prototype is available in http://www.lvc.ele.pucrio.br/projects/Liver3D/3dliver.avi.
It was also developed an environment where the user can specify different GA configurations, enable and disable search of parameters, and define their default values and search ranges. It is possible to define the training set, and fitness aggregation type (mean of fitness values, or minimum fitness value).
An evaluation of the optimal GA configuration values for the optimization of the segmentation parameters goes beyond the scope of this paper, so the GA was configured experimentally, and its convergence was verified visually. The GA configuration values are the following:

Number of generations = 30;

Population Size = 30;

Initial crossover rate = 0.8;

Final crossover rate = 0.65;

Initial mutation rate = 0.1;

Final mutation rate = 0.8;

Initial Steady State rate = 0.8;

Final Steady State rate = 0.2;

Number of sequenced experiments = 2;

Rate of seed from the first experiment to the second = 0.1;
 1.
Minimum RMS: 0.0209
 2.
G _{ L }: 0.3859
 3.
G _{ H }: 0.2809
 4.
β: 5.1929
 5.
γ: 217.414
Liver segmentation results
Evaluation  Best  Worst  Mean 

VOE  5.45  12.07  7.35 
(Score)  (78.70)  (52.82)  (71.29) 
RVD  0.63  8.12  2.19 
(Score)  (96.63)  (56.80)  (82.27) 
ASD  0.76  3.57  1.35 
(Score)  (80.85)  (10.70)  (66.25) 
RMSSD  1.69  8.22  3.05 
(Score)  (76.46)  (0)  (58.58) 
MSD  17.03  55.09  26.81 
(Score)  (77.59)  (27.21)  (64.72) 
Total Score  82.05  29.57  68.62 
The result of another exam with low score (46.18) has a single major error, caused by a peripheral nodule not classified as liver. This is again due to the adopted heuristic that models the liver tissue by a single Gaussian. As the nodule in this exam appears much darker than the liver parenchyma, its voxel intensities lie outside the range [T _{ L } ,T _{ H }] defined by the Gaussian fit. As the nodule is at the periphery, the morphological hole filling filter was not able to correct the error, and therefore the nodule region was not included in the final result. Important metrics such as RMSSD and MSD, which take into account the maximum differences between the reference and the resulting surfaces, generated low scores.
In the last exam with poor performance, we see a leak towards the mediastinum, with overall score of 45.58. This leak towards the mediastinum is commonly observed in many different approaches, mainly because the mediastinum share some intensity values with the liver and is anatomically close to the liver in a wide extent. This makes it hard to define properly the boundary between them in some cases. Some post processing can be done to avoid it, as the mediastinum has a known relative position to the liver. Another possibility would be the inclusion of the advection term to create a barrier between the liver and the mediastinum. The operator using a proper user interface could manually define this.
Our results can be compared with the results of other approaches tested upon the data and metrics compliant with the liver segmentation competition held in the Sliver07 conference. If compared with other automatic and semiautomatic methods presented in the conference's website, all of them using the same dataset used in this work, our method had a good performance, being ranked as the 6th best method among 26 (2008).
The estimation of parameters may take a long time (about one week), but it has to be executed just once, and the values are used for all exams from then on. Once the parameters have been estimated, the method is quite fast. The computation time depends on the exam size: a 210 slices exam, takes 204 seconds to segment the liver, while a 110 slices exam takes 105 seconds, on a AMD Athlon 3.0 GHz with 2GB RAM.
Analysis of sensibility
It is possible to see that Minimum RMS values range from 0.01 to 0.03 without losing much performance. G_{L} and G_{H} range from 0.3 to 0.5 and from 0.2 to 0.5 respectively with good performances. The mean curvature weight ranges satisfactorily from near 0 to 1000. It is interesting to observe the behaviour of propagation weight; it seems to have a constant asymptotic behaviour when getting far from zero value (we tested with values of the order of tens of thousands). A high propagation weight value would probably affect the finetuning of level sets near to the liver border, but since our propagation image values are very low there, it seems not to have a great impact on performance.
Nodules and Vessels Segmentation
Identification of Portal and Hepatic veins
Couinaud Liver segmentation
Conclusions
We have presented a methodology to segment the liver and its internal structures. We also propose a method to separate the liver into segments according to the Couinaud anatomical model.
Firstly the liver was segmented using a method based on level sets and an evolutionary method to estimate its parameters. The use of level sets led to a good accuracy in the tested exams, and the evolutionary approach based on GAs to estimate segmentation parameters, has been shown robust in our experiments. The GA based method for model adaptation managed to capture the inherent subjectivity present in the visual segmentation of organs in computer tomography.
The method for nodules and vessels segmentation performed well for most CT exams of the available dataset. Yet, some limitations became apparent, especially when there are nodules or veins close to the liver periphery, and also in the presence of nodules with volume similar to the liver parenchyma. These cases indicate directions for the continuation of this research aiming at a further refinement of the proposed method. The identification of portal and hepatic veins also generated results consistent anatomically.
Finally, the geometric approach to fit planes to the hepatic and portal veins managed to define Couinaud regions. The quality assessment was also visual, and the specialist consulted to evaluate the results found some inconsistencies with other anatomic models present in the literature, though the results were consistent with the Couinaud model. This result brings up a discussion about the convergence of the different existing theoretical anatomic models for the liver [29–31].
The authors believe that the method to segment the liver may be applied with some adjustments to segment other organs, especially when the distribution of voxel intensities is nearly Gaussian shaped. The evolutionary approach can be also used to estimate segmentation parameters for other organs, if a proper training set can be provided.
Declarations
Acknowledgements
The authors would like to thank FAPERJ for funding.
Authors’ Affiliations
References
 Masutani Y, Uozumi K, Masaaki A, Kuni O: Liver CT image processing: A short introduction of the technical elements. European Journal of Radiology 2006, 58: 246–251. 10.1016/j.ejrad.2005.11.044View ArticleGoogle Scholar
 Lamecker H, Zachow S, Haberl H, Stiller M: Medical Applications for Statistical 3D Shape Models. In Book Computer Aided Surgery around the Head. Edited by: S. Weber. Berlin: VDI; 2005:61. vol. 17. pp. 61Google Scholar
 Soler L, Delingette H, Malandain G, Montagnat J, Ayache N, Clement JM, Koehl C, Dourthe O, Mutter D, Marescaux J: Fully automatic anatomical, pathological, and functional segmentation from CT scans for hepatic surgery. Computer Aided Surgery 2001, 6: 131–142. 10.3109/10929080109145999View ArticleGoogle Scholar
 Fujimoto H, Gu L, Kaneko T: Recognition of Abdominal Organs Using 3D Mathematical Morphology. System and Computers in Japan 2002, 33: 75–83.View ArticleGoogle Scholar
 Kim S, Yoo S, Kim S, Kim J, Park J: Segmentation of kidney without using contrast medium on abdominal CT image. In Book 5th International Conference on Signal Processing Proceedings. Volume 2. Beijing; 2000:1147–1152. pp. 1147–1152Google Scholar
 Osher S, Sethian J: Fronts propagating with curvaturedependent speed: algorithms based on HamiltonJacobi formulations. Journal of Computational Physics 1998, 79: 12–49.MathSciNetView ArticleGoogle Scholar
 Gamarendi JF, Malpica N, Martel J, Schiavi E: Automatic Segmentation of the Liver in CT Using Level Sets Without Edges. Pattern Recognition and Image Analysis Lecture Notes in Computer Science 2007, 4477: 161–168.View ArticleGoogle Scholar
 Chan TF, Vese LA: Active Countours Without Edges. IEEE Transactions on Image Processing 2001, 10: 266–277. 10.1109/83.902291View ArticleGoogle Scholar
 FernandezdeManuel L, Rubio JL, LedesmaCarbayo MJ, Pascau J, Tellado JM, Ramon E, Desco M, Santos A: 3D liver segmentation in preoperative CT images using a levelsets active surface method. In International Conference of the IEEE EMBS. Volume 1. Minessota; 2009:3625–3628. pp. 3625–3628Google Scholar
 Oliveira DAB, Feitosa RQ, Correia MM: Liver Segmentation using Level Sets and Genetic Algorithms. In Fourth International Conference on Computer Vision Theory and Applications. Volume 2. Porto; 2009:154–159. pp. 154–159Google Scholar
 Oliveira DAB, Feitosa RQ, Correia MM: Automatic Couinaud liver and veins segmentation from CT images. In Biosignals  International Conference on BioInspired Systems and Signal Processing. Volume 1. Funchal; 2008:249–252. pp. 249–252Google Scholar
 Oliveira DAB, Feitosa RQ, Correia MM: Genetic Adaptation of Level Sets Parameters for Medical Imaging Segmentation. In Biomedical Image Analysis and Machine Learning Technologies: Applications and Techniques. Volume 1. Edited by: Gonzalez FA, Romero E. New York; 2010:150–165. pp. 150–165View ArticleGoogle Scholar
 Couinaud C: Le foie: études anatomiques et chirurgicales. Paris: Masson; 1957.Google Scholar
 Sethian J: Level Set Methods and Fast Marching Methods: Evolving Interfaces in Computational Geometry, Fluid Mechanics, Computer Vision, and Materials Science. New York: Cambridge University Press; 1999.Google Scholar
 Malladi R, Sethian JA, Vemuri BC: Shape modeling with front propagation: a level set approach. IEEE Transactions on Pattern Analysis and Machine Intelligence 1995, 17: 158–175. 10.1109/34.368173View ArticleGoogle Scholar
 McInerney T, Terzopoulos D: Deformable Models in Medical Image Analysis: A survey. Medical Image Analysis 1996, 2: 91–108.View ArticleGoogle Scholar
 Zhang Y: A survey on evaluation methods for image segmentation. Pattern Recognition 1996, 29: 1335–1346. 10.1016/00313203(95)001697View ArticleGoogle Scholar
 Chabrier S, Emile B, Rosenberger C, Laurent H: Unsupervised performance evaluation of image segmentation. EURASIP Journal on Applied Signal Processing 2006, 2006: 1–12.MathSciNetView ArticleGoogle Scholar
 Kubassuva O, Boesen M, Bliddal H: General Framework for Unsupervised Evaluation of Quality of Segmentation Results. In ICIP Proceedings. San Diego; 2008:3036–3039. pp. 3036–3039Google Scholar
 Schmiedle F, Drechsler N, Grosse D, Drechsler R: Heuristic learning based on genetic programming. Genetic Programming and Evolvable Machines 2002, 3: 363–388. 10.1023/A:1020988925923View ArticleGoogle Scholar
 Michalewicz Z: Genetic Algorithms + Data Structures = Evolution Programs. Berlin: SpringerVerlag; 1994.View ArticleGoogle Scholar
 Davis L: Handbook of Genetic Algorithms. New York: Van Nostrand Reinhold Company; 1990.Google Scholar
 Costa GAOP, Feitosa RQ, Cazes TB, Feijó B: Genetic Adaptation of Segmentation Parameters. In Book ObjectBased Image Analysis: Spatial concepts for knowledgedriven remote sensing applications. Edited by: T. Blaschke, S. Lang, G. Hayeds. New York: SpringerVerlag; 2008:679–695. pp. 679–695View ArticleGoogle Scholar
 Levenberg K: A Method for the Solution of Certain NonLinear Problems in Least Squares. The Quarterly of Applied Mathematics 1944, 2: 164–168.MathSciNetGoogle Scholar
 Marquadt D: An Algorithm for LeastSquares Estimation of Nonlinear Parameters. Journal of Applied Mathematics 1963, 11: 431–441.Google Scholar
 Soille P: Morphological Image Analysis: Principles and Applications. Heildelberg: SpringerVerlag; 2003.Google Scholar
 Heimann T, Ginneken Bv, Styner M, Arzhaeva Y, Aurich V, Bauer C, Beck A, Becker C, Beichel R, Bekes G, et al.: Comparison and Evaluation of Methods for Liver Segmentation from CT datasets. IEEE Transactions on Medical Imaging 2009, 28: 1251–1265.View ArticleGoogle Scholar
 Yoo TS, Ackerman MJ, Lorensen WE, Schroeder W, Chalana V, Aylward S, Metaxes D, Whitaker R: Engineering and Algorithm Design for an Image Processing API: A Technical Report on ITK  The Insight Toolkit. In Book Medicine Meets Virtual Reality Proceedings. Edited by: J. Westwood. Amsterdam: IOS Press; 2002:586–592. pp. 586–592Google Scholar
 Fasel J, Selle D, Evertsz C, Terrier F, Peitgen H, Gailloud P: Segmental anatomy of the liver: Poor correlation with CT. Radiology 1998, 206: 151–156.View ArticleGoogle Scholar
 Fischer L, Thorn M, Neumann J, Schäobinger M, Heimann T, Grenacher L, Meinzer H, Friess H, Bäuchler M: The segments of the hepatic veins: is there a spatial correlation to the Couinaud liver segments? European Journal of Radiology 2005, 53: 245–255. 10.1016/j.ejrad.2004.02.003View ArticleGoogle Scholar
 Machado M: Segmentação hepática de Couinaud: além dos olhos do radiologista. Radiologia Brasileira 2003, 23: 3–4.View ArticleGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.