Unsupervised Anomaly Detection in Elderly Brain MRI Using a 3D Perceptual Autoencoder
Poster No:
1519
Submission Type:
Abstract Submission
Authors:
Emani Hunter1, Bistra Iordanova1, Jinghang Li1, Howard Aizenstein1,2, Minjie Wu2
Institutions:
1Department of Bioengineering, University of Pittsburgh, Pittsburgh, Pennsylvania, USA, 2Department of Psychiatry, University of Pittsburgh, Pittsburgh, Pennsylvania, USA
First Author:
Co-Author(s):
Bistra Iordanova, PhD
Department of Bioengineering, University of Pittsburgh
Pittsburgh, Pennsylvania, USA
Department of Bioengineering, University of Pittsburgh
Pittsburgh, Pennsylvania, USA
Howard Aizenstein, MD, PhD
Department of Bioengineering, University of Pittsburgh|Department of Psychiatry, University of Pittsburgh
Pittsburgh, Pennsylvania, USA|Pittsburgh, Pennsylvania, USA
Department of Bioengineering, University of Pittsburgh|Department of Psychiatry, University of Pittsburgh
Pittsburgh, Pennsylvania, USA|Pittsburgh, Pennsylvania, USA
Introduction:
Unsupervised anomaly detection (UAD) models have shown promise in leveraging the distribution of healthy brains to detect abnormalities in magnetic resonance (MR) scans (Baur et al., 2020). These models commonly use deep learning architectures with pixel-level loss functions that tend to struggle with handling highly complex data with subtle structural changes and inter-individual variability such as normal aging and Alzheimer's Disease data. The present study proposes a 3D UAD model that completes two tasks: 1) anomaly segmentation to identify obvious, visible anomalies such as white matter hyperintensities (WMH), and 2) anomaly delineation and visualization to identify subtle WMH associated changes.
Methods:
We trained, validated, and tested our model on the Biobank Innovations for Chronic Cerebrovascular Disease With ALZheimer's Disease Study (BICWALZS) (n = 713; age range (years) = 49-89). The data consists of T2-weighted fluid attenuated inversion recovery (FLAIR) 3T scans and ground truth labels of WMH. The ground truth labels serve as a gold standard for analyzing the model's performance. We split the data into two groups based on a median split of WMH count: 1) top 25% of data below the median for subjects with no WMH and 2) the bottom 25% of data above the median for subjects with WMH. The group 1 scans were used as the "normal distribution" for training and validation (train size = 155, validation size = 38). Group 2 data was used for the testing phase (test size = 96).
The deep learning architecture leverages the combination of unsupervised anomaly detectors found in literature (Baur et al., 2021; Luo et al., 2023; Shvetsova et al., 2021). The model uses a classical autoencoder based framework with a fully connected latent space and takes in 3D volumetric images. We constructed the autoencoder to have stacked 3D convolution blocks and residual blocks for the encoder-decoder components, a single encoder-decoder skip connection, and a dense latent space. A visual of the model schematics can be seen in figure 1a and 1b. Additionally, the model uses perceptual loss to reconstruct perceptive information from the normal distribution of images (Shvetsova et al., 2021; Tuluptceva et al., 2020).
Anomalies were quantified at the axial slice level using a z-score technique, inspired by the normality score technique found in previous literature (Luo et al., 2023). A heat map of the anomalies was generated to visualize the high residuals. These anomaly maps were then binarized based on z-score thresholding, with a strict criterion of z-scores above 3 being labeled obvious anomalies to compare with WMH ground truths. Z-scores below 3 were labeled as normal or subtle changes. The model's anomaly delineation performance was evaluated using the area under the receiver operating curve (AUC-ROC) metric and the area under the precision-recall curve (AUPRC) metric. A dice similarity score was computed between the ground truth labels and the binarized anomalies to evaluate the anomaly segmentation performance.
The deep learning architecture leverages the combination of unsupervised anomaly detectors found in literature (Baur et al., 2021; Luo et al., 2023; Shvetsova et al., 2021). The model uses a classical autoencoder based framework with a fully connected latent space and takes in 3D volumetric images. We constructed the autoencoder to have stacked 3D convolution blocks and residual blocks for the encoder-decoder components, a single encoder-decoder skip connection, and a dense latent space. A visual of the model schematics can be seen in figure 1a and 1b. Additionally, the model uses perceptual loss to reconstruct perceptive information from the normal distribution of images (Shvetsova et al., 2021; Tuluptceva et al., 2020).
Anomalies were quantified at the axial slice level using a z-score technique, inspired by the normality score technique found in previous literature (Luo et al., 2023). A heat map of the anomalies was generated to visualize the high residuals. These anomaly maps were then binarized based on z-score thresholding, with a strict criterion of z-scores above 3 being labeled obvious anomalies to compare with WMH ground truths. Z-scores below 3 were labeled as normal or subtle changes. The model's anomaly delineation performance was evaluated using the area under the receiver operating curve (AUC-ROC) metric and the area under the precision-recall curve (AUPRC) metric. A dice similarity score was computed between the ground truth labels and the binarized anomalies to evaluate the anomaly segmentation performance.
Results:
Figure 1a illustrates the reconstruction performance and visualization of anomalies detected in a heatmap and binary representation. The z-scores and normality scores of brain MRI axial slices can be seen in Figure 2a and 2b. A high z-score and low normality score show up when the slice has an anomaly. Figure 2c and 2d show the ROC curve, and precision-recall curve. The area under the precision recall curve was 0.896, the area under the ROC curve was 0.889, and the dice score for anomaly segmentation of WMH was 0.403.
Conclusions:
This work develops a specialized unsupervised anomaly detection model with application to white matter and associated changes. Our work has a broader impact in complex imaging data and represents a promising tool for effectively detecting Alzheimer's Disease related changes in the brain.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s)
Lifespan Development:
Aging
Modeling and Analysis Methods:
Methods Development 1
Other Methods 2
Keywords:
Aging
Machine Learning
MRI
White Matter
Other - Deep Learning
1|2Indicates the priority used for review
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?
Are you Internal Review Board (IRB) certified? Please note: Failure to have IRB, if applicable will lead to automatic rejection of abstract.
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. Baur, C. (2021). Modeling Healthy Anatomy with Artificial Intelligence for Unsupervised Anomaly Detection in Brain MRI. Radiology: Artificial Intelligence, 3(3), e190169.
3. Luo, G. (2023). Unsupervised anomaly detection in brain MRI: Learning abstract distribution from massive healthy brains. Computers in Biology and Medicine, 154, 106610.
4. Shvetsova, N. (2021). Anomaly Detection in Medical Imaging With Deep Perceptual Autoencoders. IEEE Access, 9, 118571–118583.
5. Tuluptceva, N. (2020). Perceptual Image Anomaly Detection. Lecture Notes in Computer Science, 164–178.