Automated Localization of Fetal Organs in MRI Using Random Forests with Steerable Features K. Keraudren 1 , B. Kainz 1 , O. Oktay 1 , V. Kyriakopoulou 2 , M. Rutherford 2 , J.V. Hajnal 2 , D. Rueckert 1 1 Biomedical Image Analysis Group, Imperial College London, 2 Centre for the Developing Brain, King’s College London Size normalization Due to fetal motion, fetal MRI is typically ac- quired as stacks of 2D slices of real-time MRI, freezing in-plane motion. Motion correction methods can subsequently be applied to cor- rect the misalignment between slices and pro- vide consistent 3D data 1 . Such methods perform slice-to-volume registration and require the fetal anatomy to be isolated from surrounding mater- nal tissues. We thus propose a method to auto- matically localize the fetal heart, lungs and liver. u 0 u 0 w 0 w 0 v 0 v 0 brain brain heart heart Sagittal plane v 0 v 0 w 0 w 0 u 0 u 0 heart heart right right lung lung Transverse plane Average image of the training data after size normalization. To reduce the variability due to fetal develop- ment, we normalize the size of all fetuses by re- sampling the images to an isotropic voxel size s ga that is a function of the gestational age, so that a fetus of 30 weeks is resampled to a voxel size s 30 : s ga = CRL ga /CRL 30 × s 30 where CRL denotes the crown-rump length. Organ localization pipeline Inspired by Hough Forests 2 , the proposed method performs classification (a,c) and regression (b,d) steps using Random Forests in order to assign voxels to a certain organ then vote for the location of the organ center. A set of organ candidates is generated, then scored based on their relative position. (a) (b) (c) (d) Proposed pipeline for the automated localization of fetal organs in MRI. The center of the brain 3 is first used to steer features when detecting the heart, which then fixes an axis when detecting the lungs and liver. Knowing the location of the brain, the search for the heart only needs to explore the image region contained between two spheres. The search for the lungs and liver can similarly be restricted to a sphere around the heart. Steerable features In order to cope with the unknown orientation of the fetus, image features are extracted in a local coordinate system. At training time, the coordinate system ( ~u 0 ,~v 0 ,~ w 0 ) is defined by land- marks on the fetal anatomy. At test time, the coordinate system ( ~u,~v,~ w ) is estimated as or- gans are detected: first the brain, which fixes a point, then the heart, which fixes an axis, and finally the liver and both lungs. u u v v u u v v u u v v At test time, features are steered toward the center of the brain. Results The method was evaluated on two datasets of T2 MRI, a first dataset without motion artifacts and a second with artifacts, with gestational ages (GA) ranging from 20 to 38 weeks. Thanks to the size normalization, the same trained detector can be used across all GA. A similar performance of the detector across GA was observed. In 90% of cases, the detected heart center is within 10mm of the ground truth, which suggests that the proposed method could provide an automated initialization for the motion correction of the chest 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 when ranking organ candidates. Left lung Right lung Heart Liver 0 5 10 15 20 25 30 35 40 Distance error (mm) 1 st dataset: healthy Left lung Right lung Heart Liver 0 5 10 15 20 25 30 35 40 1 st dataset: IUGR Left lung Right lung Heart Liver Brain 0 5 10 15 20 25 30 35 40 2 nd dataset Distance error between the predicted organ centers and their ground truth for the first dataset (30 healthy and 25 IUGR fetuses) and the second dataset (64 healthy fetuses). References [1] M. Kuklisova-Murgasova, G. Quaghebeur, M. Rutherford, J. Hajnal, and J. Schnabel, “Reconstruction of Fetal Brain MRI with Inten- sity Matching and Complete Outlier Removal,” Medical Image Analysis, 2012. [2] J. Gall and V. Lempitsky, “Class-specific Hough Forests for Object Detection,” in CVPR, 2009. [3] K. Keraudren, V. Kyriakopoulou, M. Ruther- ford, J. V. Hajnal, and D. Rueckert, “Localisa- tion of the Brain in Fetal MRI Using Bundled SIFT Features,” in MICCAI, 2013. [4] B. Kainz, C. Malamateniou, M. Murgasova, K. Keraudren, M. Rutherford, J. V. Hajnal, and D. Rueckert, “Motion Corrected 3D Reconstruc- tion of the Fetal Thorax from Prenatal MRI,” in MICCAI, 2014. Conclusion & Future work Acquisition Motion correction Sagittal Coronal Transverse We presented a pipeline which, in combination with automated brain detection 3 , enables the au- tomated localization of the lungs, heart and liver in fetal MRI. The localization results can be used to initialize a segmentation or motion correction, and to orient the 3D volume with respect to the fetal anatomy to facilitate clinical diagnosis. Preliminary work used the rough segmentation produced by the detection process to generate a mask for the fetal trunk, with morphological op- erations and a region growing algorithm. Future work will focus on a slice-by-slice segmentation in order to increase the quality of the motion cor- rected volume.