Unsupervised anomaly detection and segmentation in brain MRI: a comparative study

1423 

Abstract Submission 

Junhyeok Lee¹, Won Hee Lee¹

¹Kyung Hee University, Yongin

Junhyeok Lee  
Kyung Hee University
Yongin

Won Hee Lee  
Kyung Hee University
Yongin

Unsupervised anomaly detection (UAD) plays a crucial role in identifying anomalies by learning the distribution of normal data without relying on labeled anomalies. These methods are widely used in various industries, including manufacturing damage detection, financial fraud detection, cyber intrusion, and medical diagnosis (Chandola et al., 2009). A previous study showed the effectiveness of UAD algorithms in detecting chronic infarction, suggesting their potential for brain MRI applications (van Hespen et al., 2021). However, limited research exists on the performance of UAD algorithms in the context of brain tumor detection and segmentation (Baur et al., 2021). To address this gap, this study aims to conduct a comprehensive evaluation of various UAD algorithms specifically for brain tumor detection and segmentation.

We used T2-weighted MRI data from two different datasets: 272 patients diagnosed with glioma from the multimodal brain tumor segmentation challenge 2020 (BraTS2020; age range 19 – 87 years) and 252 healthy individuals from the Cambridge Centre for Ageing and Neuroscience (Cam-CAN; 126 females; age range 18 – 88 years). We compared the performance of seven UAD methods on anomaly detection and segmentation. These methods included memory bank approach CFA (Lee et al., 2022), the reconstruction-based method DRAEM (Zavrtanik et al., 2021), the patch distribution-based method PaDiM (Defard et al., 2021), the normalizing flow-based methods CS-Flow (Rudolph et al., 2022), and FastFlow (Yu et al., 2021), and the knowledge distillation-based methods STFPM (Wang et al., 2021), and RD4AD (Deng et al., 2021). To assess the performance of these UAD methods, we used the area under the receiver operator curve (AUC) at image-level for anomaly detection and pixel-level for segmentation.

Figure 1 displays the performance of various UAD methods for anomaly detection and segmentation in brain MRI. The normalizing flow-based models (CS-Flow and FastFlow) achieved the highest detection performance (image-level AUC = 99.47 – 99.78). However, these methods showed the lowest segmentation performance (pixel-level AUC = 81.47 – 83.18) and led to an increased rate of false positives in segmentation tasks (Figure 2a). STFPM and DRAEM, while having the lower detection performance (image-level AUC = 87.51 – 88.55; pixel-level AUC = 96.35 – 96.43), had better segmentation performance compared to the flow-based models. FastFlow, CS-Flow, DRAEM, STFPM showed a substantial performance gap between detection and segmentation, with differences ranging from 7.80 to 18.31. In contrast, RD4AD and CFA exhibited a small difference between detection and segmentation performance ranging from 1.82 to 2.06. RD4AD and CFA achieved superior segmentation performance (pixel-level AUC = 97.67 – 98.18). However, we found that RD4AD and STFPM had a tendency to misidentify the lateral ventricle areas in normal samples as abnormal regions (Figure 2b).

Our comparative evaluation of various UAD methods for brain MRI reveal their potential applicability. Our results indicate that UAD methods can be adapted for brain anomaly detection and segmentation, while emphasizing the need for optimal model tuning and refinement to improve the performance of the UAD methods. Moreover, the experimental results indicate the potential for improved UAD performance by improving the model's understanding of normal anatomical structures with similar intensities to abnormal regions in MRI scans. Incorporating anatomical information into the learning process may be a key strategy for achieving enhanced performance, offering more reliable support in the clinical diagnosis and monitoring of brain pathologies.

Classification and Predictive Modeling ¹

Segmentation and Parcellation ²

Machine Learning

MRI

STRUCTURAL MRI

Other - Anomaly Detection