TL;DR
This paper introduces DiffuCE, a novel diffusion model for enhancing low-quality Cone-Beam Computed Tomography (CBCT) images, crucial for improving the accuracy and safety of radiotherapy. DiffuCE addresses limitations of existing methods by using latent space alignment to bridge domain gaps, enabling effective enhancement without paired training data and achieving expert-level performance.
What Problem Does This Paper Address?
Cone-Beam Computed Tomography (CBCT) is widely used in Image-Guided Radiotherapy (IGRT) due to its speed and low radiation dose. However, CBCT images often suffer from poor quality, characterized by noise and blurred details, making it difficult for radiologists to accurately identify anatomical structures. This compromises the precision of IGRT. Existing enhancement methods often rely on specific imaging settings or assumptions about noise distributions, limiting their generalizability. Furthermore, many diffusion model approaches for medical image enhancement are computationally expensive, requiring extensive training and long inference times, which hinders their practical clinical deployment.
Methods
DiffuCE is built upon a denoising diffusion probabilistic model (DDPM) backbone. Its key innovations include:
- Latent Alignment: A "Domain Bridging Encoder" (DBE) is introduced to bridge the distribution gap between CBCT and CT images in the latent space. This is achieved by training a projector that maps noisy CBCT latent embeddings to a noisy CT distribution, effectively aligning the domains. This allows the model to learn from diverse data without requiring perfectly paired CT and CBCT scans.
- Conditional Diffusion Denoiser (CDD): This component enhances image quality by removing noise while preserving details. It leverages "constraints preprocessing" (e.g., bone, lucent areas, low-frequency features extracted via wavelet transform) as conditional inputs to guide the denoising process. The use of LoRA (Low-Rank Adaptation) modules helps fine-tune the pre-trained latent diffusion model efficiently.
- Conditional Refinement Decoder (CRD): To ensure pixel-level accuracy, the CRD integrates conditions into the reconstructed latent embedding to control the decoding process. It employs an "Adaptive DualScope Loss" (ADL) that combines region-wise loss (focusing on specific areas like bone and lucent regions), perceptual loss (evaluating overall image similarity), and adversarial loss (using a discriminator for GAN-like training) to achieve high-quality reconstruction.
Datasets
DiffuCE was evaluated on two datasets:
- SynthRAD2023 CBCT-to-CT Pelvis Dataset: A pair-wise dataset containing 180 3D CBCT and CT image pairs collected from three different centers. This dataset was split into training and validation sets.
- Private Dataset: A dataset from a collaborated medical center, comprising 4177 2D CT images and 2816 2D CBCT Chest images, collected over approximately one month. This dataset was used for training the CDD and DBE, and a subset was used for qualitative research.
Key Insights and Ideas
- Bridging Domain Gaps in Latent Space is Crucial: The "Distribution Gap" between CBCT and CT images is a significant challenge. DiffuCE's latent alignment strategy effectively addresses this, enabling better generalization and reducing the need for perfectly paired data.
- Conditional Guidance Enhances Detail Preservation: Using specific clinical features (bone, lucent areas, low-frequency components) as conditions significantly improves the denoising process, preventing the generation of incorrect patterns and preserving fine details.
- Pixel-Level Refinement is Essential for Clinical Utility: While latent space operations are efficient, the CRD's focus on pixel-level accuracy is vital for clinical applications where every detail matters. The adaptive loss function balances different aspects of reconstruction quality.
- Fine-tuning Pre-trained Models is Viable for Medical Imaging: The paper demonstrates that fine-tuning a foundation diffusion model (using LoRA) is an efficient strategy for medical image enhancement, making advanced techniques more accessible.
Key Results
- Private Dataset: DiffuCE achieved the best scores in LPIPS (human visual perception similarity) and FID (data sample distribution), indicating superior perceptual quality and distribution matching. It secured second place in conventional metrics like PSNR and SSIM.
- SynthRAD2023 Dataset: DiffuCE outperformed or matched several state-of-the-art methods, demonstrating competitive performance across different datasets.
- Expert Assessments: Five radiologists rated DiffuCE as the best among various methods, highlighting its clinical relevance and expert-level performance.
- Ablation Studies: Removing the Domain Bridging Encoder (DBE) led to a slight improvement in metrics but a loss of contrast. Removing the Conditional Diffusion Denoiser (CDD) resulted in distorted soft tissue representation. Removing the Conditional Refinement Decoder (CRD) caused a significant drop in most metrics, underscoring its importance.
Limitations and Future Work
- Limitations:
- The framework's inference speed is slower than GAN-based algorithms, which could be a hurdle for real-world deployment.
- Without conditional control, preserving intricate details can be challenging, leading to distortions and lower performance on conventional pixel-level evaluations.
- Future Work:
- The authors suggest adding more condition-specific modules (e.g., for muscle, soft tissue, or tumors) to further improve detail preservation.
- The latent alignment mechanism offers potential for broader domain adaptation in diffusion models, suggesting future research in this area.
Summary updated