"Neighboring Slice Noise2Noise: Self-Supervised Medical Image Denoising from Single Noisy Image Volume"

When medical image slices team up, they clean up their own noise.

Clean up noisy medical scans by comparing slice neighbors - no clean data needed.

This paper introduces NS-N2N (Neighboring Slice Noise2Noise), a self-supervised medical image denoising method that works with single noisy image volumes. It leverages spatial continuity between neighboring slices to achieve high-quality denoising without requiring paired clean-noisy training data.

So NS-N2N teaches neural nets to denoise medical images by learning from their neighbors.

-----

https://arxiv.org/abs/2411.10831

🔍 Original Problem:

Current medical image denoising methods either need paired clean-noisy images for training or assume pixel-wise independent noise. Both assumptions fail in real medical settings, making existing solutions impractical.

-----

🛠️ Solution in this Paper:

→ NS-N2N uses neighboring slices within a single noisy image volume to construct weighted training data

→ It identifies matched regions between adjacent slices where tissue structures remain continuous

→ The method applies strong low-pass filtering to compute weight matrices that distinguish matched from unmatched regions

→ Training combines regional consistency loss and inter-slice continuity loss for optimization

→ The entire process operates purely in the image domain, avoiding device-specific reconstruction issues

-----

💡 Key Insights:

→ Medical image volumes have natural spatial continuity between slices

→ Noise between neighboring slices is independent, enabling Noise2Noise training

→ Device-independent image-domain processing increases clinical applicability

-----

📊 Results:

→ Outperformed state-of-the-art self-supervised methods in both synthetic and real-world tests

→ Required fewer training epochs compared to existing methods

→ Achieved higher PSNR and SSIM scores across different noise levels