Multimodal MRI-Based Unsupervised Brain Tumor Detection Using Modality Translation and Anomaly Discrimination: A...

Accurate brain tumor detection in magnetic resonance imaging (MRI) is essential for clinical diagnosis and treatment planning. Annotating large datasets of each tumor type for supervised models is costly and time-consuming. Therefore, this study investigated unsupervised tumor detection with only healthy samples by leveraging multimodal MRI characteristics.

577 brain MRIs from healthy adults are used for training. 3788 multicenter MRIs encompassing three tumor types are used for validation. A modality translating network (MTN) and an anomaly discriminating network (ADN) were trained with only healthy T1WI and T2WI. The MTN was trained to translate healthy T1WI into healthy T2WI. The ADN learned to compare healthy T2WI with abnormal T2WI, which is manually synthesized from healthy T1WI and T2WI, to detect the synthetic abnormal regions. During validation, the MTN translated tumor-containing T1WI into T2WI that appeared free of lesions. The ADN then compared the translated lesion-free T2WI with the original tumor-containing T2WI to detect and segment tumor regions.

Our proposed model achieved an average precision (AP) of 77% and a dice similarity coefficient (DSC) of 54%, outperforming state-of-the-art unsupervised methods 0.17 in AP and 0.06 in DSC.

The proposed model only needs healthy MRI samples for training, which reduces the burden of manual annotating huge amounts of different kinds of brain tumors. The detection performance across multiple centers and tumor types demonstrated the generalizability of the proposed approach. Leveraging multimodal MRI characteristics can improve brain tumor detection with higher accuracy and better generalization.