Multi-Modal Brain Tumor Data Completion Based on Reconstruction Consistency Loss.

Area of Science:

• Medical imaging
• Artificial intelligence
• Neuroscience

Background:

• Multi-modal brain MRI is crucial for tumor segmentation, but data corruption poses challenges.
• Existing deep learning methods struggle with inter-slice semantic and intra-slice structural information for image completion.
• Image completion is vital for pre-processing, aiding diagnosis and reducing patient burden.

Purpose of the Study:

• To propose a novel generative adversarial network (GAN) framework, RAGAN, for completing missing multi-modal brain MRI data (T1, T1ce, FLAIR) from T2 data.
• To enhance the accuracy and efficiency of multi-modal brain MRI data completion.
• To improve subsequent brain tumor segmentation and diagnosis using completed MRI data.

Main Methods:

• Developed a novel generative adversarial network (GAN) framework named RAGAN.
• Generator utilizes T2 modal images and multi-modal classification labels for cyclic supervised training to restore arbitrary modal images.
• Discriminator features a multi-branch network to assess similarity of generated images and their essential visual features to target modalities.

Main Results:

• RAGAN effectively reconstructs brightness, resolution, location, and morphology of brain tissue across different MRI modalities.
• Segmentation validation using UNet achieved 77.58% accuracy; using RES_UNet with depth supervision yielded 88.76% accuracy.
• The proposed method achieved a 2% higher DICE value compared to the state-of-the-art TC-MGAN algorithm for data completion.

Conclusions:

• RAGAN demonstrates significant potential for completing missing multi-modal brain MRI data, outperforming existing methods in specific metrics.
• The enhanced image quality and improved segmentation accuracy facilitate more reliable brain tumor diagnosis.
• This approach offers a cost-effective and less burdensome solution for patients undergoing MRI examinations.