Unsupervised anomaly detection in brain MRI: Learning abstract distribution from massive healthy brains
Guoting Luo1, Wei Xie2, Ronghui Gao2
1Huaxi MR Research Center (HMRRC), Department of Radiology, West China Hospital of Sichuan University, Chengdu, China.
Computers in Biology and Medicine
|January 28, 2023
Summary
This study introduces an unsupervised deep learning method for detecting brain abnormalities like glioblastoma using only normal MRI scans. The approach shows promise for computer-aided diagnosis in radiology.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neurology
Background:
- Accurate detection of brain abnormalities is crucial for timely diagnosis and treatment.
- Existing methods often require large datasets of labeled abnormal samples, limiting their applicability.
- Unsupervised learning offers a potential solution for detecting novel or rare abnormalities.
Purpose of the Study:
- To develop a general unsupervised anomaly detection method for brain MR images.
- To automatically detect various brain abnormalities without training on abnormal samples.
Main Methods:
- A novel three-dimensional deep autoencoder network was proposed.
- The model was trained and validated on 578 normal T2w MR brain volumes.
- Evaluation was performed on datasets containing glioblastoma, multiple sclerosis, and cerebral infarction.
Main Results:
- The method achieved high AUCs for anomaly detection: 0.844 (glioblastoma), 0.858 (multiple sclerosis), and 0.807 (cerebral infarction).
- Mean Dice score for glioblastoma segmentation was 0.462.
- The network can generate anomaly heatmaps for visualization.
Conclusions:
- The unsupervised method successfully detected diverse brain anomalies.
- This approach enables detection of arbitrary brain abnormalities without labeled data.
- It holds potential as an automated tool to support radiological diagnostic workflows.
