Automated Localization of Fetal Organs in MRI Using Random Forests with Steerable Features (Miccai 2015)

Automated Localization of Fetal Organs in MRI
Using Random Forests with Steerable Features
Kevin Keraudren1
, Bernhard Kainz1
, Ozan Oktay1
, Vanessa Kyriakopoulou2
,
Mary Rutherford2
, Joseph V. Hajnal2
and Daniel Rueckert1
1
Biomedical Image Analysis Group, Imperial College London
2
Department Biomedical Engineering, King’s College London
Abstract. Fetal MRI is an invaluable diagnostic tool complementary to
ultrasound thanks to its high contrast and resolution. Motion artifacts
and the arbitrary orientation of the fetus are two main challenges of fetal
MRI. In this paper, we propose a method based on Random Forests with
steerable features to automatically localize the heart, lungs and liver in
fetal MRI. During training, all MR images are mapped into a standard
coordinate system that is defined by landmarks on the fetal anatomy
and normalized for fetal age. Image features are then extracted in this
coordinate system. During testing, features are computed for different
orientations with a search space constrained by previously detected land-
marks. The method was tested on healthy fetuses as well as fetuses with
intrauterine growth restriction (IUGR) from 20 to 38 weeks of gestation.
The detection rate was above 90% for all organs of healthy fetuses in the
absence of motion artifacts. In the presence of motion, the detection rate
was 83% for the heart, 78% for the lungs and 67% for the liver. Growth
restriction did not decrease the performance of the heart detection but
had an impact on the detection of the lungs and liver. The proposed
method can be used to initialize subsequent processing steps such as
segmentation or motion correction, as well as automatically orient the
3D volume based on the fetal anatomy to facilitate clinical examination.
1 Introduction
Although the main focus of fetal Magnetic Resonance Imaging (MRI) is in imag-
ing of the brain due to its crucial role in fetal development, there is a growing
interest in using MRI to study other aspects of fetal growth, such as assessing the
severity of intrauterine growth restriction (IUGR) [4]. In this paper, we present
a method to automatically localize fetal organs in MRI (heart, lungs and liver),
which can provide automated initialization for subsequent processing steps such
as segmentation or motion correction [8].
The main challenge for the automated localization of fetal organs is the
unpredictable orientation of the fetus. Two methodological approaches can be
identified from the current literature on the localization of the fetal brain in MRI.
The first approach seeks a detector invariant to the fetal orientation [6,9,7]. The
second approach learns to detect fetuses in a known orientation and rotates

2 K. Keraudren et al.
Fig. 1: Overview of the proposed pipeline for the automated localization of fetal
organs in MRI.
the test image [1]. We adopt this second approach in the present paper, but
as illustrated by steerable features [12], it is more efficient to extract features
in a rotating coordinate system than to rotate the whole 3D volume. We thus
propose to extend the Random Forests (RF) framework for organ localization
[10] with the extraction of features in a local coordinate system specific to the
anatomy of the fetus.
A RF classifier [3] is an ensemble method for machine learning which averages
the results of decision trees trained on random subsets of the training dataset,
a process called bagging. Similarly, a RF regressor is an ensemble of regression
trees. A RF classifier assigns a label to each voxel, resulting in a segmentation of
the input image while a RF regressor assigns an offset vector to each voxel, allow-
ing it to vote for the location of landmarks. Classification and regression forests
can be combined so that only voxels assigned to a certain class can vote for the
location of specific landmarks. This is of particular relevance when detecting
landmarks within the fetus whose position has little correlation with surround-
ing maternal tissues. While [5] proposed to form trees that randomly alternate
decision and regression nodes (Hough Forest), we chose to sequentially apply a
classification and a regression forest. This enables us to benefit from generic RF
implementations [11] while focusing on the key step of feature extraction.
At training time, image features are learnt in a coordinate system that is
normalized for the age of the fetus and defined by the anatomy of the fetus
(Fig. 2). At test time, assuming that the center of the brain is known, image
features are extracted in a coordinate system rotating around the brain in order
to find the heart. The lungs and liver are then localized using a coordinate
system in which one axis is parallel to the heart-brain axis. An overview of our
proposed pipeline is presented in Fig. 1. In the remainder of this paper, we will
present our proposed method in Section 2 and evaluate it in Section 3 on two
datasets of MRI scans. The first dataset, used for training and leave-one-out
cross validation, consists of scans of 30 healthy and 25 IUGR subjects without
significant motion artifacts. The second dataset (64 subjects) is used to evaluate
the performance of the proposed method in the presence of motion artifacts.
2 Method

Automated Localization of Fetal Organs in MRI 3
u0u0
v0v0
w0w0
liverliver
Coronal plane
u0u0
w0w0
v0v0
brainbrain
heartheart
Sagittal plane
v0v0
w0w0
u0u0
heartheart
rightright
lunglung
Transverse plane
Fig. 2: Average image of 30 healthy fetuses after resizing (Section 2) and align-
ment defined by anatomical landmarks (Section 2). It highlights the character-
istic intensity patterns in fetal MRI for the brain, heart, lungs and liver.
Preprocessing: There is a large variability in the MR scans of fetuses due
to fetal development. In particular, the length of the fetus increases by 50%
between 23 weeks and term (40 weeks) [2]. We thus propose to normalize the
size of all fetuses by resampling the images to an isotropic voxel size sga that is a
function of the gestational age (GA), so that a fetus of 30 weeks is resampled to
a voxel size s30: sga = CRLga/CRL30 ×s30, where CRL denotes the crown-rump
length [2]. This differs from previously proposed methods which either ignored
the GA [1,6,7] or only used it to define size constraints [9]. Moreover, fetal MRI
is typically acquired as stacks of 2D images that freeze in-plane motion but
form a 3D volume corrupted by fetal and maternal motion. The images are thus
preprocessed using median filtering to attenuate motion artifacts.
Detecting candidate locations for the heart: In theory, a detector could be
trained on aligned images while at test time, it would be applied to all possible
rotated versions of the 3D volume. However, in practice it would be too time
consuming to explore an exhaustive set of rotations. In particular, the integral
image that is used to extract intensity features would need to be computed
for all tested rotations. Instead of rotating the image itself, we thus propose to
rotate the sampling grid of the image features. This is done by defining a local
orientation of the 3D volume for every voxel.
When training the classifier, all voxels are oriented using the anatomy of the
fetus (Fig. 2 and Fig. 3.a): u0 points from the heart to the brain, v0 from the left
lung to the right lung and w0 from back to front (w0 = u0 ×v0). When applying
the detector to a new image, in theory, all possible orientations (u, v, w) should be
explored. However, assuming that the location of the brain is already known [9],
u can be defined as pointing from the current voxel x to the brain (Fig. 3.b). In
this case, only the orientation of the plane Px,u passing through x and orthogonal
to u remains undefined. (v, w) are then randomly chosen so that they form an
orthonormal basis of Px,u. The rationale for this definition of u is that when x
is positioned at the center of the heart, Px,u is the transverse plane of the fetal
anatomy (Fig. 2). To address the sensitivity to the brain localization, a random
offset is added to the estimated brain center with a 10mm range comparable to
the prediction errors reported in [9]. Additionally, knowing the location of the

u0u0v0v0
u0u0v0v0
u0u0v0v0
uu
vv
uu
vv
uu
vv
(a) (b)
Fig. 3: (a) When training the classifier, features (light blue squares) are extracted
in the anatomical orientation of the fetus: u0 points from the heart (green) to the
brain (red), and v0 points from the left lung (blue) to the right lung (yellow). The
pink contour corresponds to the liver. (b) At testing time, features are extracted
in a coordinate system rotating around the brain.
brain, the search for the heart only needs to explore the image region contained
between two spheres of radii R1 and R2 (Fig. 4) that are independent of the GA
thanks to the size normalization.
The localization of the heart is performed in two stages: a classification forest
is first applied to obtain for each voxel x a probability p(x) of being inside the
heart (Fig 5.a). All voxels with p(x) > 0.5 are then passed on to a regression
forest, which predicts for each of them the offset to the center of the heart.
Using the predicted offsets, the probabilities are then accumulated in a voting
process, and the voxels receiving the highest sum of probabilities are considered
as candidate centers for the heart (Fig 5.b). During classification, in order to
increase the number of true positives, an initial random orientation of the plane
Px,u is chosen and the detector is successively applied by rotating (v, w) with
an angular step θ. The value of (v, w) maximizing the probability of the current
voxel to be inside the heart is selected, and only this orientation of Px,u is passed
on to the regression step.
Four types of features are used in the RFs: tests on image intensity, gradient
magnitude, gradient orientation and distance from the current voxel to the brain.
In order to be invariant to intensity shifts [10], image intensity or gradient mag-
nitude features select two axis-aligned cube patches of random sizes at random
offsets from the current voxel and compare their mean intensity. All features are
extracted in the local orientation (u, v, w). Similarly, during the voting process,
the predicted offsets are computed with the same local orientation.
Detecting candidate locations for the lungs and liver: For each candidate
location of the heart, candidate centers for the lungs and liver are detected.
Similarly to the heart detection, a classification RF first classifies each voxel
as either heart, left lung, right lung, liver or background. For each organ, a

R1R1
R2R2
R
3
R
3
Fig. 4: Knowing the location of the brain, the search for the heart can be re-
stricted to a narrow band within two spheres of radius R1 and R2 (red). Sim-
ilarly, knowing the location of the heart, the search for the lungs and the liver
can be restricted to a sphere of radius R3 (green).
regression RF is then used to vote for the organ center. During training, the
images are oriented according to the fetal anatomy (Fig. 2). During testing, both
the location of the brain and heart are known and only the rotation around the
heart-brain axis needs to be explored. For each voxel x, we can hence set u = u0,
and only the orientation of the plane Px,u0
remains unknown. (v, w) are thus
randomly chosen to form an orthonormal basis of Px,u0
.
In the set of features for the RFs, the distance from the current voxel to
the brain, which was used in the detection of the heart, is now replaced by the
distance to the heart. Additionally, the projection of the current voxel on the
heart-brain axis is used in order to characterize whether the current voxel is be-
low or above the heart relative to the brain. Similarly to the heart detection, the
detector is run with an angular step θ. For each voxel, the value of θ maximizing
the probability of being inside one of the organs of the fetus is selected.
Spatial optimization of candidate organ centers: The voting results from
the regression RF are first smoothed with a Gaussian kernel. Local maxima
are then extracted to form a set of candidates xl for the center of each organ,
l ∈ L = { heart, left lung, right lung, liver }, along with their individual scores
p(xl). Each combination of candidate centers is assigned a score (Eq. 1) based
on the individual scores p(xl) and the spatial configuration of all organs:
l∈L
λp(xl) + (1 − λ)e− 1
2 (xl−¯xl) Σ−1
l (xl−¯xl)
(1)
where ¯xl is the mean position and Σl the covariance matrix for the organ l in
the training dataset, computed in the anatomical coordinate system of the fetus.
λ is a weight between individual scores and the spatial configuration of organs.
Once the organ centers have been localized, the voting process can be reversed
in order to obtain a rough segmentation (Fig. 5.c).

(a) (b) (c)
Fig. 5: (a) Probability of a voxel to be inside the heart. (b) Accumulated regres-
sion votes for the center of the heart. (c) After all organs have been located,
the voting process can be reversed, resulting in a rough segmentation for each
organ: heart (blue), left lung (red), right lung (green), liver (yellow).
3 Experiments
Datasets: Two datasets of 1.5T T2-weighted ssTSE images were used in this
paper. The first dataset was used to train and validate the model in a leave-
one-out cross validation, in the absence of motion artifacts. It consists of 55
scans covering the whole uterus for 30 healthy subjects and 25 IUGR subjects
(scanning parameters: TR 1000ms, TE 98ms, 4mm slice thickness, 0.4mm slice
gap). The second dataset is used to test the model in the presence of motion
artifacts. There are 224 scans, from 64 healthy subjects, which do not all fully
cover the fetal body (scanning parameters: TR 15000ms, TE 160ms, 2.5mm slice
thickness, 1.5mm slice overlap). A ground truth segmentation for all fetal organs
was provided by an expert for the first dataset. Due to the large number of images
and the presence of motion artifacts, only the center of each organ has been
manually annotated in the second dataset. The location of the brain is passed
to the detector for the first dataset in order to evaluate the proposed method on
its own while the method of [9] is used to localize the brain automatically in the
second dataset, thus evaluating a fully automated localization of fetal organs.
Both datasets cover the range of gestational ages from 20 to 38 weeks.
Results: The results are reported as detection rates in Table 1, where a de-
tection is considered successful for the organ l if the predicted center is within
the ellipsoid defined by (xl − yl) Σ−1
l (xl − yl) < 1 where yl is the ground truth
center. For the first dataset, Σl is the covariance matrix of the ground truth
segmentation while for the second dataset, it is defined from training data as
in Section 2. The distance error between the predicted centers and their ground
truth are reported in Fig. 6. Although healthy and IUGR subjects were mixed
in the training data, results are reported separately to assess the performance of
the detector when fetuses are small for their gestational age. Training two detec-
tors on each subset of the data did not improve the results. The higher detection

Table 1: Detection rates for each organs.
Heart Left lung Right lung Liver
1st
dataset: healthy 90% 97% 97% 90%
1st
dataset: IUGR 92% 60% 80% 76%
2nd
dataset 83% 78% 83% 67%
Left
lung
Right
lung
Heart Liver
0
5
10
15
20
25
30
35
40
Distanceerror(mm)
1st
dataset: healthy
Left
lung
Right
lung
Heart Liver
0
5
10
15
20
25
30
35
40
1st
dataset: IUGR
Left
lung
Right
lung
Heart Liver Brain
0
5
10
15
20
25
30
35
40
2nd
dataset
Fig. 6: Distance error between the predicted organ centers and their ground truth
for the first dataset (30 healthy fetuses and 25 IUGR fetuses) and the second
dataset (64 healthy fetuses).
rate for the lungs than for the heart for the healthy subjects of the first dataset
(Table 1) is due to the detector overestimating the size of the heart (Fig 5.c),
which can be linked to the signal loss when imaging the fast beating heart. A
decrease in performance can be observed for the detection of the lungs in IUGR
fetuses, which can be associated with a reduction in lung volume reported in [4].
The asymmetry in the performance of the detector between the left and right
lungs can be explained by the presence of the liver below the right lung, leading
to a stronger geometric constraint in Eq. 1.
For the second dataset, the largest error for the detection of the brain detec-
tion using the method of [9] is 21mm, which is inside the brain and is sufficient
for the proposed method to succeed. In 90% of cases, the detected heart center
is within 10mm of the ground truth, which suggests that the proposed method
could be used to provide an automated initialization for the motion correction of
the chest [8]. Beside motion artifacts, the main difference between the first and
second dataset is that the second dataset does not fully cover the fetal abdomen
in all scans, which could explain its lower performance when detecting the liver.
Training the detector takes 2h and testing takes 15min on a machine with 24
cores and 128 GB RAM. The parameters in the experiments were chosen as
follows: s30 = 1mm, maximal patch size 30 pixels, maximal offset size 20 pixels,
λ = 0.5, θ = 90◦
.

4 Conclusion
We presented a pipeline which, in combination with automated brain detection,
enables the automated localization of the lungs, heart and liver in fetal MRI.
The localization results can be used to initialize a segmentation or motion cor-
rection, as well as to orient the 3D volume with respect to the fetal anatomy to
facilitate clinical diagnosis. The key component to our method is to sample the
image features in a local coordinate system in order to cope with the unknown
orientation of the fetus. During training, this coordinate system is defined by
the fetal anatomy while at test time, it is randomly oriented, within constraints
set by already detected organs. In future work, the possibility to merge the
classification and regression steps into a Hough Forest [5] will be investigated.
References
1. Anquez, J., Angelini, E., Bloch, I.: Automatic Segmentation of Head Structures on
Fetal MRI. In: ISBI. pp. 109–112. IEEE (2009)
2. Archie, J.G., Collins, J.S., Lebel, R.R.: Quantitative Standards for Fetal and
Neonatal Autopsy. American Journal of Clinical Pathology 126(2), 256–265 (2006)
3. Breiman, L.: Random Forests. Machine learning 45(1), 5–32 (2001)
4. Damodaram, M., Story, L., Eixarch, E., Patkee, P., Patel, A., Kumar, S., Ruther-
ford, M.: Foetal Volumetry using Magnetic Resonance Imaging in Intrauterine
Growth Restriction. Early Human Development (2012)
5. Gall, J., Lempitsky, V.: Class-specific Hough Forests for Object Detection. In:
CVPR. pp. 1022–1029. IEEE (2009)
6. Ison, M., Donner, R., Dittrich, E., Kasprian, G., Prayer, D., Langs, G.: Fully
Automated Brain Extraction and Orientation in Raw Fetal MRI. In: Workshop on
Paediatric and Perinatal Imaging, MICCAI. pp. 17–24. Springer (2012)
7. Kainz, B., Keraudren, K., Kyriakopoulou, V., Rutherford, M., Hajnal, J.V., Rueck-
ert, D.: Fast Fully Automatic Brain Detection in Fetal MRI using Dense Rotation
Invariant Image Descriptors. In: ISBI. pp. 1230–1233. IEEE (2014)
8. Kainz, B., Malamateniou, C., Murgasova, M., Keraudren, K., Rutherford, M., Ha-
jnal, J., Rueckert, D.: Motion Corrected 3D Reconstruction of the Fetal Thorax
from Prenatal MRI. In: Golland, P., Hata, N., Barillot, C., Hornegger, J., Howe, R.
(eds.) MICCAI 2014, part II, LNCS, vol. 8674, pp. 284–291. Springer, Heidelberg
(2014)
9. Keraudren, K., Kyriakopoulou, V., Rutherford, M., Hajnal, J.V., Rueckert, D.:
Localisation of the Brain in Fetal MRI Using Bundled SIFT Features. In: Mori,
K., Sakuma, I., Sato, Y., Barillot, C., Navab, N. (eds.) MICCAI 2013, part I,
LNCS, vol. 8149, pp. 582–589. Springer, Heidelberg (2013)
10. Pauly, O., Glocker, B., Criminisi, A., Mateus, D., Möller, A., Nekolla, S., Navab,
N.: Fast Multiple Organ Detection and Localization in Whole-Body MR Dixon
Sequences. In: Fichtinger, G., Martel, A., Peters, T. (eds.) MICCAI 2011, part III,
LNCS, vol. 6893, pp. 239–247. Springer, Heidelberg (2011)
11. Pedregosa, F., et al.: Scikit-learn: Machine Learning in Python. Journal of Machine
Learning Research 12, 2825–2830 (2011)
12. Zheng, Y., Barbu, A., Georgescu, B., Scheuering, M., Comaniciu, D.: Fast Au-
tomatic Heart Chamber Segmentation from 3D CT Data using Marginal Space
Learning and Steerable Features. In: ICCV. pp. 1–8. IEEE (2007)

Automated Localization of Fetal Organs in MRI Using Random Forests with Steerable Features (Miccai 2015)

More Related Content

What's hot (15)

Viewers also liked (20)

Similar to Automated Localization of Fetal Organs in MRI Using Random Forests with Steerable Features (Miccai 2015) (20)

More from Kevin Keraudren (12)

Recently uploaded (20)

Automated Localization of Fetal Organs in MRI Using Random Forests with Steerable Features (Miccai 2015)