Ano-swinMAE: Unsupervised Anomaly Detection in Brain MRI using swin Transformer based Masked Auto Encoder
Keywords: Anomaly Detection, Masked Auto Encoder, swin Transformer, Brain MRI
Abstract: The advanced deep learning-based Autoencoding techniques have enabled the introduction of efficient Unsupervised Anomaly Detection (UAD) approaches. Several autoencoder-based approaches have been used to solve UAD tasks. However, most of these approaches do not have any constraints to ensure the removal of pathological features while restoring the healthy regions in the pseudo-healthy image reconstruction. To remove the occurrence of pathological features, we propose to utilize an Autoencoder which deploys a masking strategy to reconstruct images. Additionally, the masked regions need to be meaningfully inpainted to enforce global and local consistency in the generated images which makes transformer-based masked autoencoder a potential approach. Although the transformer models can incorporate global contextual information, they are often computationally expensive and dependent on a large amount of data for training. Hence we propose to employ a Swin transformer-based Masked Autoencoder (MAE) for anomaly detection (Ano-swinMAE) in brain MRI. Our proposed method Ano-swinMAE is trained on a healthy cohort by masking a certain percentage of information from the input images. While inferring, a pathological image is given to the model and different segments of the brain MRI slice are sequentially masked and their corresponding generation is accumulated to create a map indicating potential locations of pathologies. We have quantitatively and qualitatively validated the performance increment of our method on the following publicly available datasets: BraTS (Glioma), MSLUB (Multiple Sclerosis) and White Matter Hyperintensities (WMH). We have also empirically evaluated the generalisation capability of the method in a cross modality data setup.
Submission Number: 314
Loading