Neural network fusion: a novel CT-MR Aortic Aneurysm image segmentation method

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Multimodal imaging (CT, MR, PET) enhances diagnostic accuracy by leveraging complementary strengths.
• Segmenting Abdominal Aortic Aneurysm (AAA) from CT and MR images is crucial for treatment but challenging due to registration issues.

Purpose of the Study:

• To develop an efficient deep neural network for multimodal medical image segmentation.
• To address challenges in registering non-affinely deformed CT and MR images for AAA analysis.

Main Methods:

• Proposed a novel deep neural network approach fusing high-level features of separate CT and MR segmentation networks.
• Enabled end-to-end training with non-registered images, reducing training time and resource inefficiency.
• Hypothesized and demonstrated shared neural representations for common concepts across modalities.

Main Results:

• Achieved shorter training times compared to separate network training.
• Learned a shared representation of the aorta across CT and MR modalities.
• Demonstrated that similar aneurysm conditions exhibit shorter feature-level distances in the neural network, aiding image registration.

Conclusions:

• The proposed fused network offers an efficient and effective solution for multimodal medical image segmentation.
• Shared feature representations facilitate improved image registration and analysis for conditions like AAA.
• This method holds promise for enhancing diagnostic capabilities and treatment planning in medical imaging.