Deep Generative Anomaly Detection for Structural Anomalies in Fetal Brain with Ventriculomegaly
Poster No:
1454
Submission Type:
Abstract Submission
Authors:
Sungmin You1,2, Carlos Simon Amador Izaguirre1, Seungyoon Jeong1,2, Hyukjin Yun1,2,3, Ellen Grant1,4,5, Kiho Im1,2,3
Institutions:
1Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital, Boston, MA, 2Division of Newborn Medicine, Boston Children’s Hospital, Harvard Medical School, Boston, MA, 3Department of Pediatrics, Harvard Medical School, Boston, MA, 4Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital, Harvard Medical School, Boston, MA, 5Department of Radiology, Boston Children’s Hospital, Harvard Medical School, Boston, MA
First Author:
Sungmin You
Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital|Division of Newborn Medicine, Boston Children’s Hospital, Harvard Medical School
Boston, MA|Boston, MA
Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital|Division of Newborn Medicine, Boston Children’s Hospital, Harvard Medical School
Boston, MA|Boston, MA
Co-Author(s):
Carlos Simon Amador Izaguirre
Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital
Boston, MA
Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital
Boston, MA
Seungyoon Jeong
Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital|Division of Newborn Medicine, Boston Children’s Hospital, Harvard Medical School
Boston, MA|Boston, MA
Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital|Division of Newborn Medicine, Boston Children’s Hospital, Harvard Medical School
Boston, MA|Boston, MA
Hyukjin Yun
Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital|Division of Newborn Medicine, Boston Children’s Hospital, Harvard Medical School|Department of Pediatrics, Harvard Medical School
Boston, MA|Boston, MA|Boston, MA
Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital|Division of Newborn Medicine, Boston Children’s Hospital, Harvard Medical School|Department of Pediatrics, Harvard Medical School
Boston, MA|Boston, MA|Boston, MA
Ellen Grant
Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital|Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital, Harvard Medical School|Department of Radiology, Boston Children’s Hospital, Harvard Medical School
Boston, MA|Boston, MA|Boston, MA
Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital|Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital, Harvard Medical School|Department of Radiology, Boston Children’s Hospital, Harvard Medical School
Boston, MA|Boston, MA|Boston, MA
Kiho Im
Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital|Division of Newborn Medicine, Boston Children’s Hospital, Harvard Medical School|Department of Pediatrics, Harvard Medical School
Boston, MA|Boston, MA|Boston, MA
Fetal Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital|Division of Newborn Medicine, Boston Children’s Hospital, Harvard Medical School|Department of Pediatrics, Harvard Medical School
Boston, MA|Boston, MA|Boston, MA
Introduction:
Fetal ventriculomegaly (VM) means the enlargement of the cerebral ventricles diagnosed in utero and occurs in up to 2 per 1000 births. (Pisapia, Sinha, Zarnow, Johnson, & Heuer, 2017) The potential effects of VM on neurodevelopment can be variable depending on the severity from little impact to developmental delays, motor skill challenges, and intellectual disabilities.(Chervenak et al., 1984; Weichert et al., 2010) Early diagnosis of VM using fetal MRI is essential for the better identification of etiology and guidance of prognosis. (Pisapia et al., 2017) This study proposes a deep generative anomaly detection model for the diagnosis of VM based on structural anomalies in fetal brain MRI.
Methods:
This study was approved by the local Institutional Review Board at Boston Children's Hospital. For the training and test of the proposed model, 151 typically developing (TD) fetuses (gestational weeks [GW]: 31.3±4.0, range: 22.0-38.7; sex: 69/49/28 [male/female/unknown]) and 46 fetuses with VM (GW: 29.8±4.5, range: 20.2-38.0; sex: 33/11/2) were included in this study. The TD subjects were divided into training-testing groups, which resulted in 121 subjects for the training set and 30 subjects for the test set while all VM subjects were used only for the test.
We used our pipeline for fetal MRI processing which has been validated in several studies (Yun et al., 2022) including brain masking, non-uniformity correction, and slice-to-volume registration (Kuklisova-Murgasova, Quaghebeur, Rutherford, Hajnal, & Schnabel, 2012) to preprocess the fetal MRIs. We extracted 30 center slices on each view for training and testing after cropping them to different sizes according to their view, sagittal 158×126, coronal 110×126, and axial 110×158. We normalize the intensities of images with min-max normalization.
Our anomaly detection model is based on the variational autoencoder (VAE) (Kingma & Welling, 2013) composed of four convolutional blocks on the encoder and decoder. [Figure 1] The proposed model was trained for each view during 2000 epochs with the Adam optimizer with a learning rate of 1e-4 and the mean squared error loss. For the evaluation, the averaged pixel-wise mean absolute error between the input and reconstruction was used to compute the anomaly score for each image. We performed the area under the receiver operating characteristic(AUROC) analysis to confirm the feasibility of classification between TD and VM. We extracted the center slice as a representative image of a volume and performed the Mann-Whitney U-test to compare the distribution of anomaly scores between groups.
We used our pipeline for fetal MRI processing which has been validated in several studies (Yun et al., 2022) including brain masking, non-uniformity correction, and slice-to-volume registration (Kuklisova-Murgasova, Quaghebeur, Rutherford, Hajnal, & Schnabel, 2012) to preprocess the fetal MRIs. We extracted 30 center slices on each view for training and testing after cropping them to different sizes according to their view, sagittal 158×126, coronal 110×126, and axial 110×158. We normalize the intensities of images with min-max normalization.
Our anomaly detection model is based on the variational autoencoder (VAE) (Kingma & Welling, 2013) composed of four convolutional blocks on the encoder and decoder. [Figure 1] The proposed model was trained for each view during 2000 epochs with the Adam optimizer with a learning rate of 1e-4 and the mean squared error loss. For the evaluation, the averaged pixel-wise mean absolute error between the input and reconstruction was used to compute the anomaly score for each image. We performed the area under the receiver operating characteristic(AUROC) analysis to confirm the feasibility of classification between TD and VM. We extracted the center slice as a representative image of a volume and performed the Mann-Whitney U-test to compare the distribution of anomaly scores between groups.
·[Figure 1. Overview of the variational autoencoder for anomaly detection in fetal MRI]
Results:
Our proposed model showed 0.836 from the AUROC analysis on the test set using the sagittal view. Furthermore, the Mann-Whitney showed statistically significant higher anomaly scores (p<0.001) in the VM than in TD with the sagittal view, where an elevation of anomaly scores distribution in the VM than the TD is observed. [Figure2]
·[Figure 2. (a) AUROC analysis for the sagittal view model; (b) Comparison of anomaly score distribution from the sagittal slide between TD and VM]
Conclusions:
Our VAE-based anomaly detection model for diagnosing fetal VM on MRI showed an AUROC of 0.836 in the sagittal view. This demonstrates the potential of the anomaly detection model for the accurate prenatal detection of VM, which can aid in understanding VM's etiology and improving prognosis guidance. For a future study, including other views and developing a framework for aggregating slice-wise results into volume-level predictions will further enhance the diagnostic power. This model can also be used for detecting and diagnosing other various neurodevelopmental disorders in utero.
Disorders of the Nervous System:
Neurodevelopmental/ Early Life (eg. ADHD, autism) 2
Lifespan Development:
Normal Brain Development: Fetus to Adolescence
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Methods Development
Keywords:
Data analysis
Development
MRI
STRUCTURAL MRI
Other - deep generative model
1|2Indicates the priority used for review
Provide references using author date format
Kingma, D. P., & Welling, M. (2013). Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114.
Kuklisova-Murgasova, M., Quaghebeur, G., Rutherford, M. A., Hajnal, J. V., & Schnabel, J. A. (2012). Reconstruction of fetal brain MRI with intensity matching and complete outlier removal. Med Image Anal, 16(8), 1550-1564. doi:10.1016/j.media.2012.07.004
Pisapia, J. M., Sinha, S., Zarnow, D. M., Johnson, M. P., & Heuer, G. G. (2017). Fetal ventriculomegaly: diagnosis, treatment, and future directions. Child's Nervous System, 33, 1113-1123.
Weichert, J., Hartge, D., Krapp, M., Germer, U., Gembruch, U., & Axt-Fliedner, R. (2010). Prevalence, characteristics and perinatal outcome of fetal ventriculomegaly in 29,000 pregnancies followed at a single institution. Fetal diagnosis and therapy, 27(3), 142-148.
Yun, H. J., Lee, H. J., Lee, J. Y., Tarui, T., Rollins, C. K., Ortinau, C. M., . . . Im, K. (2022). Quantification of sulcal emergence timing and its variability in early fetal life: Hemispheric asymmetry and sex difference. Neuroimage, 263, 119629. doi:10.1016/j.neuroimage.2022.119629