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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate brain tumor detection via MRI is crucial for diagnosis and treatment.
• Supervised models require costly and time-consuming annotation of large tumor datasets.
• Unsupervised methods offer a potential solution to reduce annotation burden.

Purpose of the Study:

• To develop an unsupervised brain tumor detection method using only healthy MRI samples.
• To leverage multimodal MRI characteristics for improved detection accuracy.
• To reduce the reliance on extensive manual annotation for training AI models.

Main Methods:

• A modality translating network (MTN) and anomaly discriminating network (ADN) were trained on healthy T1-weighted imaging (T1WI) and T2-weighted imaging (T2WI).
• The MTN translates healthy T1WI to T2WI; the ADN detects anomalies by comparing synthesized and real T2WI.
• During validation, the model identifies tumors by detecting discrepancies between translated lesion-free T2WI and original tumor-containing T2WI.

Main Results:

• The proposed unsupervised model achieved an average precision (AP) of 77% and a Dice similarity coefficient (DSC) of 54%.
• Performance significantly outperformed state-of-the-art unsupervised methods by 0.17 in AP and 0.06 in DSC.
• The model demonstrated robust detection across multiple centers and tumor types.

Conclusions:

• The developed model effectively detects brain tumors using only healthy MRI data, significantly reducing annotation efforts.
• The approach shows strong generalizability across different centers and tumor types.
• Utilizing multimodal MRI data enhances brain tumor detection accuracy and model generalization.