Robust Region Growing Bad
Intrathoracic Airway Tree Segmentation
Rˆo mulo Pinho,Sten Luyckx,and Jan Sijbers
University of Antwerp,Physics Department,VisionLab,Belgium
{romulo.pinho;jan.sijbers}@ua.ac.be
{sten.luyckx}@student.ua.ac.betribute词根
Abstract.In this paper,we prent a mi-automatic region growing
xsysinfo
algorithm to gment the intrathoracic airway tree from3-d CT images.
A common problem with region growing is leakage.In order to limit
leakage,our method bounds the gmentation using cylinders of adaptive
orientation and dimensions.The leaks are detected bad on anatomical
information of the airways and an algorithm to avoid them is propod.
We also prent an algorithm to automatically lect a ed point for the
gmentation.The method was tested on a datat of40patients and
results were quantitatively evaluated bad on ground truth data.
1Introduction
jp是什么意思Airway tree gmentation is the process of identifying and extracting from vol-umetric medical images the structures of the respiratory system that lead the air into the lungs.With the result of the gmentation,doctors and rearchers can make measurements,check for abnormalities and generally be assisted in diagnosing dias in the respiratory system.
In this work,we concentrate on the gmentation of the lower airway tree, namely the trachea,bronchi and bronchioli.Due to the natural complexity of the airways,with veral branching levels,and noi or other artefacts prent in the image,the gmentation is far from trivial.A common method to solve the problem is region growing[1],and mi-and fully-automated region growing algorithms have been ud to gment the airways[2–4].In this process,the ur provides o
ne or more ed points inside the airway structure.From the points,a region is grown by recursively aggregating voxels that pass a certain test of similarity.Common similarity tests check differences in intensity between neighbouring voxels.
One common problem of region growing algorithms is leakage.In the ca of the airway tree gmentation,a thin wall parates the structure from neigh-bouring organs and air inside the lungs.Noi or other artefacts can create holes in this wall and,since the airway lumen and the lung interior have similar voxel intensities,the entire lung can be aggregated to the region.Another problem specific to airway tree gmentation is the early collap of branches.In this ca,the growing process stops too early,resulting in only partially gmented branches.
Tschirren et al.propod an algorithm that takes advantage of the fact that the airway tree is a hierarchical combination of cylindrically shaped objects[5]. In their algorithm,cylinders of adaptive radius and orientation bound the g-mentation,facilitating the process of leak detection.Later,Pinho et al.propod improvements to[5]and introduced new ways of detecting leaks using anatom-ical,instead of pure image features[6].In the prent work,we build upon[6] and add the following contributions:
–propo a heuristic algorithm to automatically lect a ed point inside the trachea,since chest CT scans often include the upper airways and other regions,complicating this task;
–u cylinders of adaptive height as well as adaptive radius in order to bound the gmentation;
–propo a new strategy to avoid leaks,by taking into account the fact that they grow through small holes on the edges of the structure being gmented.
The propod method was evaluated with a datat of40patients,subdivided into training and testing groups.Measures of number of branches,airway tree length,and leakage were taken in order to evaluate the method,by comparing it to ground truth data.
2Methodwash
We begin with a review of the method propod in[5]and the ideas introduced in[6].In[5],multieded fuzzy connectivity(MFC)[7]was ud to gment the airways’walls and lumen in an iterative process which places adaptive cylinders (or ROIs)around the region to be grown.Airway walls and lumen compete for voxels bad on an affinity valueψ∈[0,1]assigned as a function of voxel intensities.The ROIs bound the region growing and t limits to possible leaks.
A leak detector assumes that leaks have a“spongy”structure,with many holes and tunnels.Once a leak is detected with a special morphological operator,the algorithm goes back to the previous step and repeats the gmentation,using what the authors called“directional affinity”.This strategy avoids new leaks by assigning affinity values as a function of the intensity of a voxel and of its neighbours lying in the direction of the ROI.Airway branching is detected by computing the skeleton of a region within an ROI,using distance transforms.The branches of the skeleton,their spatial orientation,and the interctions between the region and the borders of the ROI determine the radius and orientation of the ROIs of the next step.The heights of ROIs may change if the gmentation stops exactly at a branching point.This process continues until no more voxels are aggregated.
In[6],the authors propod improvements to the above algorithm.First,ex-ecution speed was incread with a simplification of the skeleton computation: instead of computing the real skeleton,an approximation was obtained by di-rectly linking the centres of gravity of interctions between a region and their respective ROI.For interctions occurring on the side and upper borders of
Fig.1.Avoiding leaks.The gmentation is repeated with an increasing neighbourhood mask until no leaks are detected.
the ROIs,a global centre of gravity for the ROI was obtained and connected to
the regions of interction.This approximation is certainly not preci enough
with respect to skeleton accuracy,but suffices for the estimation of the radii and
orientations of the ROIs of the next step.
The cond improvement dealt with the detection of leaks.Instead of the
purely image bad approach adopted in[5],the leak detection us anatomical
作品集培训
knowledge about the airways.For instance,the number of offspring branches
from one level to the next is usually not larger than3or4and the radius of
a branch is normally a decreasing function of its length and branching level.
By checking the number of branches and their areas resulting from interctions
with ROIs,leaks can be easily detected.Similar ideas were ud in[8],with
wavefront propagation algorithms,and in[9],with region growing.
In order to further improve[6]by detecting more airway branches and reach-
ing deeper into the lungs,we propo to extend the region growing with two new
ideas.Thefirst is to u ROIs of adaptive height instead of only adaptive radius,
such that h l=Hρ(l−1)and h l≥H min,where l≥1is the current branching level,h l is the height of the associated cylinder,in millimetres,H is the default
initial height,ρ∈[0,1]is the height change ratio,and H min is the minimum
height limit.By doing this,short branches at higher branching levels,which
otherwi would not interct an ROI,can be detected.
Secondly,whenever the gmentation within an ROI is repeated due to a
leak,each candidate voxel and its neighbourhood within a mask are analyd.
Only if the voxel and all its non-visited neighbours pass the similarity test(voxel
湖南美容学校intensity lower than a threshold T in our ca)is the voxel aggregated to the
region.Each time the gmentation is repeated,a mask of higher radius is ud,
until no leaks are detected or a maximum number of attempts is reached.The
reasoning behind this approach is that a leak always occurs due to the prence
of holes on the boundaries of the region being gmented,so we are basically
trying to discover their sizes.In contrast to the direction affinity adopted in[5],
our neighbour affinity technique allows more possibilities of continuing with the
gmentation,while trying to avoid leaks.The2-d scheme of Fig.1illustrates
the idea and Algorithm1details the process.
Algorithm1region grow(image,ed)
1:/*intialization and computation of1st ROI*/
2:while¬empty(roi queue)do
3:r←pop(roi queue)
4:roi region grow(image,r)
5:has leak←detect leak(r){using anatomical information}
6:if has leak∧(count leak<max count leak)then
7:t neighbour affinity(mask[count leak]){t mask to avoid leak}
8:count leak←count leak+1
9:ret(image,r)
10:push(roi queue,r)
11:el
12:if has leak then
13:remove leaking branches(r){may remove all branches}
14:end if
15:roi list←process roi(r)
16:for all r i∈roi list do
17:push(roi queue,r i)
18:end for
19:count leak←0
20:t neighbour affinity(null){switch neighbour affinity off}
21:end if
22:end while
In the algorithm,Step1compris a quence of steps to compute thefirst ROI,using the given image and ed point,and push it onto an ROI-queue. After growing a region within the ROI at the front of the queue,Step5detects leaks using anatomical information.A leak is identified when the ROI splits into more than5regions for levels1through4and into more than3regions for higher levels.In addition,a leak is also identified when the area from a branch to its children increas by a factor f≥2.The regions of interction between the ROI and the region grown corresponding to leaks are put in a list.If a leak is detected,Step7switches neighbour affinity on by providing the next neighbourhood mask to be ud in the similarity tests.The ROI is ret and pushed back onto the queue in Steps9and10,respectively.If leaks are still prent after trying all masks,Step13removes from the list obtained in Step 5the corresponding branches.Step15process the remaining branches of the current ROI and returns a list of ROIs for the next iteration,which are pushed onto the queue.Finally,Step20switches neighbour affinity offand the process restarts.
Seed Point Selection As mentioned previously,region growing algorithms need one or more ed points to mark the start of the gmentation.Algorithms that automate the ed point lection for the gmentation of the airways typi-cally detect a circular region near the centre of a slice of the image
volume.The region is suppod to correspond to the trachea,and the ed point is taken as
Fig.2.A situation where other structures may mislead the detection of the trachea in an axial slice of the CT scan.
its centre of gravity.This process may fail if the chon slice contains misleading ,if the CT scan contains parts of the upper airways)or does not contain the trachea at all.Fig.2illustrates the former ca with a slice contain-ing the trachea,the oesophagus,and a tumour.In the prent work,we propo a more robust method to automatically lect a ed point inside the trachea.
For one axial slice N of the image volume,the method works as in Algorithm2.Let us u the threshold below operation to turn all voxels with intensity below a certain threshold to white and the rest to black[10].Step1 thusfinds the best threshold to gment the air in the image,which includes the areas inside the lungs and airway lumen,using Otsu’s method[11].Step2 applies a masked,morphological closing operation to the slice in order tofill all holes.Step3identifies8-connected regions in the resulting image and labels them.Step6removes ,all regions with size s≤S min pixels.Step10 takes care of eliminating narrowed ,with excentricity e>e max,and steps12through15identify the region of the slice with highest excentricity,R e
i
.
After the steps,a number of regions may be left in each slice.The regions compri the trachea and areas corresponding to air outside the lungs,the upper airways,the lungs,etc.The challenge is then to choo the slice containing only the trachea or at least to correctly identify it when other structures are prent. For this,we minimi a function of veral parameters,in order to favour:
–slices with fewer regions,since,in general,the upper part of the trachea tends to appear alone in the image;
–slices that maximi e,of R e
i ,given that the upper part of the trachea,just
below the infraglotic cavity,tends to be elliptical;
–slices in which the major axis of R e
i is aligned with the sagittal plane;
Algorithm 2find trachea(slice i )
1:
threshold(slice i ){gment air }2:
clo(slice i )3:
regions i ←label(slice i )4:
guesswhatfor all R j ∈regions i do 5:
if size(R j )<S min then 6:
remove(regions i ,R j ){remove noi }7:
el 8:
e ←excentricity(R j )9:
if e >e max then 10:
adultery
remove(regions i ,R j ){remove narrow regions }11:
el 12:
if e >max e then 13:
gobblemax e ←e 14:
R e i ←R j 15:
end if 16:
end if 17:
end if 18:end for
–slices with lower indexes,since the arch is for the top of the trachea (as-
suming slice 0coincides with the top position of the CT scan);
–slices in which R e i is small,which avoids confusion with the lungs;–slices in which R e i maximis the area of the ellip,so that only “regularly”shaped ellips are chon.
We therefore define the minimisation as arg min i =1..N f (i,e i ,a i ,s i ,r i )=n r i i N +s i S
+(1−e i )+a i +(1−r i ) ,(1)where {e i ,a i ,r i ∈[0,1]}.In this equation,i is the slice index,n r i ≥1is the
number of regions of the slice,s i is the size of R e i in pixels,with S being a max-imum size threshold,e i is the excentricity of R e i ,a i is the angle between R e i ’s major axis and the sagittal direction,and r i is a measure of area maximisation.The latter is computed by taking the ratio between the number of pixels of R e i and the area of its corresponding ellip.Lastly,the lected ed point is the centre of gravity of the R e i that minimis f (·).
平面设计版面
3Experiments
As stated in Section 1,the propod method was tested with a datat of 40patients,provided as part of the workshop and airway gmentation chal-lenge EXACT09:Extraction of Airways from CT 2009.The data was subdi-vided into one training and one testing group,each with 20patients,CASE40,respectively.The gmen-tation was evaluated by a team of trained obrvers.The aim of the workshop
