Pure Vision Transformer (CT-ViT) with Noise2Neighbors Interpolation for Low-Dose CT Image Denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Convolutional Neural Networks (CNNs) face challenges in medical image processing, including limited feature representation and computational costs.
• Transformers offer a promising alternative to address these CNN limitations in medical image analysis.
• Accurate patient diagnosis relies on preserving fine spatial details in medical images.

Purpose of the Study:

• To introduce a pure Vision Transformer (ViT) based denoising model for low-dose computed tomography (LDCT) image processing.
• To evaluate the efficacy of the ViT model against CNN-based and hybrid CNN-ViT models for LDCT denoising.
• To assess the model's ability to preserve high and low-frequency information and fine structural details.

Main Methods:

• A U-Net framework incorporating Vision Transformer (ViT) modules and Noise2Neighbor (N2N) interpolation was developed.
• The proposed pure ViT model was trained and tested on five diverse datasets of low-dose and normal-dose CT image pairs.
• Quantitative (SSIM, PSNR) and visual comparisons were conducted against established CNN and hybrid models.

Main Results:

• The pure ViT model demonstrated a 15-20% increase in Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR) compared to pure CNN models.
• Self-attention mechanisms in transformers outperformed traditional CNNs in quantitative metrics.
• Visual analysis confirmed superior performance of the ViT model in reconstructing fine structural details in CT images.

Conclusions:

• Pure Vision Transformers effectively address limitations of CNNs in medical image denoising.
• The proposed ViT-based U-Net model offers significant improvements in LDCT image quality and diagnostic detail preservation.
• This research highlights the potential of transformers for advancing medical image processing techniques.