Unsupervised Anomaly Detection for Fetal Brain MRI using Two-Stage Denoising Autoencoder (ω-DAE）
Poster No:
1563
Submission Type:
Abstract Submission
Authors:
Yingqi Hao1, Mingxuan Liu2, Juncheng Zhu2, Hongjia Yang2, Yi Liao3, Haibo Qu3, Qiyuan Tian2,4
Institutions:
1Weixian College, Tsinghua University, Beijing, China, 2School of Biomedical Engineering, Tsinghua University, Beijing, China, 3Department of Radiology, West China Second University Hospital, Sichuan University, Chengdu, China, 4Tsinghua Laboratory of Brain and Intelligence, Beijing, China
First Author:
Co-Author(s):
Yi Liao
Department of Radiology, West China Second University Hospital, Sichuan University
Chengdu, China
Department of Radiology, West China Second University Hospital, Sichuan University
Chengdu, China
Haibo Qu
Department of Radiology, West China Second University Hospital, Sichuan University
Chengdu, China
Department of Radiology, West China Second University Hospital, Sichuan University
Chengdu, China
Qiyuan Tian
School of Biomedical Engineering, Tsinghua University|Tsinghua Laboratory of Brain and Intelligence
Beijing, China|Beijing, China
School of Biomedical Engineering, Tsinghua University|Tsinghua Laboratory of Brain and Intelligence
Beijing, China|Beijing, China
Introduction:
Fetal germinal matrix and intraventricular hemorrhage (GMH-IVH) is the most common type of fetal brain hemorrhage and holds significant clinical importance. MRI is pivotal for the early detection of subtle prenatal GMH-IVH lesions, which manifest as hypointense signals in T2-weighted MRI (Manganaro et al., 2023). Deep learning emerges as a powerful tool for detecting such lesions from MRI, but they require large datasets from diseased subjects with ground-truth labels for training. Unfortunately, such datasets are difficult to obtain due to the rarity of GMH-IVH (0.5‰–0.9‰) and the challenges in annotating fetal MRI (Dunbal et al., 2021). Moreover, the low signal-to-noise ratio (SNR) in fetal brain MRI causes existing unsupervised anomaly detection methods like DAE (Kascenas et al., 2022), skip-TS (Liu et al., 2023), ADFA (Huang et al., 2023) to struggle with precise lesion segmentation. To address these challenges, we propose the ω-DAE based on 3D convolutional neural networks (CNNs) for unsupervised diagnosis of GMH-IVH.
Methods:
Data Acquisition. The study used MRI data from 462 pregnant women, with 436 normal fetal brains and 26 brains with GMH-IVH lesions. 2D T2-weighted TSE image data were acquired in axial, coronal, and sagittal directions. The NeSVoR (Xu et al., 2023) method was used for slice-to-volume motion correction and reconstruction of a single super-resolved fetal brain volume. Experienced radiologists manually annotated the hemorrhage regions.
Network Architecture (Fig. 1) The proposed ω-DAE for unsupervised GMH-IVH diagnosis consists of two stacked autoencoders. The first is a Variational Autoencoder (VAE) (Zimmerer et al., 2019; Zhou et al., 2020) for initial denoising to improve the SNR of fetal brain volume, while the second is a Denoising Autoencoder (DAE) for unsupervised anomaly detection (Kascenas et al., 2022). The training is divided into two stages, utilizing only normal brain data. In training stage I, the DAE component of the ω-DAE is frozen while the VAE is trained to reconstruct the input fetal MRI volumes for initial denoising. In training stage II, the pre-trained VAE is frozen, and the noise map upsampled from 16×16×16 Gaussian noise is introduced to train the DAE component. During the testing phase, the heatmap for anomaly detection is generated by subtracting the intermediate output of ω-DAE from the final output, which allows for precise visualization of the hemorrhagic areas.
Network Architecture (Fig. 1) The proposed ω-DAE for unsupervised GMH-IVH diagnosis consists of two stacked autoencoders. The first is a Variational Autoencoder (VAE) (Zimmerer et al., 2019; Zhou et al., 2020) for initial denoising to improve the SNR of fetal brain volume, while the second is a Denoising Autoencoder (DAE) for unsupervised anomaly detection (Kascenas et al., 2022). The training is divided into two stages, utilizing only normal brain data. In training stage I, the DAE component of the ω-DAE is frozen while the VAE is trained to reconstruct the input fetal MRI volumes for initial denoising. In training stage II, the pre-trained VAE is frozen, and the noise map upsampled from 16×16×16 Gaussian noise is introduced to train the DAE component. During the testing phase, the heatmap for anomaly detection is generated by subtracting the intermediate output of ω-DAE from the final output, which allows for precise visualization of the hemorrhagic areas.
Results:
Data of 390 normal brains were used for training and data of 72 brains were used for testing (46 normal brains and 26 brains with GMH-IVH). The proposed ω-DAE was compared to a vanilla DAE in unsupervised diagnosis of GMH-IVH in terms of segmentation accuracy.
Figure 2 presents the results of these models in axial, coronal, and sagittal orientations. Both the ω-DAE and DAE heatmaps demonstrate a high intensity in the hemorrhagic regions. However, the heatmap from the DAE model contains noise from irrelevant areas, whereas the heatmap from the ω-DAE focuses more on the hemorrhagic regions.
Quantitative analysis indicates that the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPR) for the DAE model at the case level for classification are 0.8874 and 0.8474, respectively, while those for the ω-DAE model are 0.8893 and 0.8591. Furthermore, the AUROC at the pixel level for segmentation is 0.9404 for DAE and 0.9495 for ω-DAE. The superior performance of ω-DAE is attributed to its two-stage denoising process, which mitigates the inherent noise in fetal brain MRI, thereby enhancing unsupervised anomaly detection.
Figure 2 presents the results of these models in axial, coronal, and sagittal orientations. Both the ω-DAE and DAE heatmaps demonstrate a high intensity in the hemorrhagic regions. However, the heatmap from the DAE model contains noise from irrelevant areas, whereas the heatmap from the ω-DAE focuses more on the hemorrhagic regions.
Quantitative analysis indicates that the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPR) for the DAE model at the case level for classification are 0.8874 and 0.8474, respectively, while those for the ω-DAE model are 0.8893 and 0.8591. Furthermore, the AUROC at the pixel level for segmentation is 0.9404 for DAE and 0.9495 for ω-DAE. The superior performance of ω-DAE is attributed to its two-stage denoising process, which mitigates the inherent noise in fetal brain MRI, thereby enhancing unsupervised anomaly detection.
Conclusions:
We present an innovative unsupervised method for fetal brain MRI anomaly detection using the two-stage denoising autoencoder (ω-DAE) and prove its effectiveness in detecting GMH-IVH of fetal brain.
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Methods Development 1
Keywords:
Modeling
MRI
STRUCTURAL MRI
Other - Fetal brain
1|2Indicates the priority used for review
·Figure 1. Proposed ω-DAE Model and its pipeline. The ω-DAE consists of a stacked VAE and DAE. The kernel size of average pooling is 2×2 with the stride of 2. Noise is added for the training stage.
·Figure 2. Results on the Test Dataset. Each case is demonstrated in axial, coronal, and sagittal directions for the ω-DAE and DAE models. In contrast to ω-DAE, DAE results contain more noises.
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
2. Zhou, L., Deng, W., & Wu, X. (2020). Unsupervised anomaly localization using VAE and beta-VAE. arXiv preprint arXiv:2005.10686.
3. Zimmerer, D., Kohl, S. A., Petersen, J., Isensee, F., & Maier-Hein, K. H. (2018). Context-encoding variational autoencoder for unsupervised anomaly detection. arXiv preprint arXiv:1812.05941.
4. Manganaro, L., Capuani, S., Gennarini, M., Miceli, V., Ninkova, R., Balba, I., Galea, N., Cupertino, A., Maiuro, A., Ercolani, G., & Catalano, C. (2023). Fetal MRI: What’s new? A short review. European Radiology Experimental, 7(1), Article 1. https://doi.org/10.1186/s41747-023-00358-5
5. Dunbar, M. J., Woodward, K., Leijser, L. M., & Kirton, A. (2021). Antenatal diagnosis of fetal intraventricular hemorrhage: systematic review and meta‐analysis. Developmental Medicine & Child Neurology, 63(2), 144-155.
6. Huang, Y., Liu, G., Luo, Y., & Yang, G. (2023, October). Adfa: Attention-augmented differentiable top-k feature adaptation for unsupervised medical anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP) (pp. 206-210). IEEE.
7. Xu, J., Moyer, D., Gagoski, B., Iglesias, J. E., Grant, P. E., Golland, P., & Adalsteinsson, E. (2023). NeSVoR: implicit neural representation for slice-to-volume reconstruction in MRI. IEEE transactions on medical imaging, 42(6), 1707-1719.
8. Liu, M., Jiao, Y., & Chen, H. (2023, May). Skip-st: Anomaly detection for medical images using student-teacher network with skip connections. In 2023 IEEE International Symposium on Circuits and Systems (ISCAS) (pp. 1-5). IEEE.