HARU-Net: Hybrid Attention Residual U-Net for Edge-Preserving Denoising in Cone-Beam Computed Tomography

This paper introduces HARU-Net, a novel Hybrid Attention Residual U-Net architecture that integrates hybrid attention transformers and residual learning to effectively denoise low-dose Cone-Beam Computed Tomography (CBCT) images while preserving critical anatomical edges, outperforming state-of-the-art methods in both image quality metrics and computational efficiency.

The Gist

Imagine you are trying to take a beautiful, high-definition photo of a delicate, intricate sculpture (like a human jawbone) in a dark room. To protect the sculpture from the harsh light, you have to use a very dim flashlight. The problem? The dim light makes the photo look grainy and fuzzy, like an old, static-filled TV screen. You can see the general shape, but the fine details—the tiny cracks, the smooth curves, the sharp edges—are lost in the "snow."

This is exactly the problem doctors face with CBCT scans (a type of 3D X-ray used in dentistry and ENT). To keep radiation doses low and safe for patients, the machines use "dim light," which results in noisy, grainy images. Doctors need these images to be crystal clear to spot tiny problems like root canals or small fractures, but the noise often hides them.

The Old Way vs. The New Way

The Old Way (Classical Methods):
Think of trying to clean a muddy window with a wet rag. If you scrub too hard to get the mud off, you smear the picture and lose the details. If you scrub too gently, the mud stays. Traditional software struggles to find that perfect balance; it either leaves the noise or blurs the important edges.

The "Deep Learning" Way (Previous Attempts):
Scientists tried using Artificial Intelligence (AI) to fix this. Imagine hiring a super-smart art restorer who has seen millions of clean paintings. This AI can guess what the clean picture should look like. However, there's a catch: to train this AI, you need pairs of photos—one dirty and one perfectly clean of the exact same object. In medicine, you can't take a second, high-radiation photo of a patient just to get a "clean" reference; that would hurt them. So, we didn't have enough "clean" training data.

Enter HARU-Net: The "Smart Hybrid" Restorer

The authors of this paper created a new AI called HARU-Net. Think of it as a master restorer who uses a hybrid toolkit to fix the grainy CBCT photos.

Here is how HARU-Net works, using simple analogies:

1. The Training Ground (The Cadaver Dataset)

Since they couldn't scan living patients twice, they used a clever workaround. They scanned 21 human jawbones from a cadaver lab using a very high-quality, high-dose setting (the "clean" photos). Then, they used a computer to simulate the grainy noise on these clean images. Now they had perfect pairs: a clean version and a noisy version of the same jawbone. This taught the AI exactly how to turn noise back into clarity.

2. The Architecture: A Team of Specialists

HARU-Net isn't just one big brain; it's a team of specialists working together, organized like a U-shape (hence the "U-Net" name).

• The Encoder (The Detective): As the AI looks at the noisy image, it breaks it down into smaller and smaller pieces, trying to understand the "big picture" of the noise and the anatomy.
• The Bottleneck (The Brain): This is the deepest part of the network. Here, HARU-Net uses a special Residual Hybrid Attention Group (RHAG).
  • Analogy: Imagine a detective who has seen the whole crime scene. Instead of just looking at one clue, this detective connects the dots across the entire room to understand the context. This helps the AI understand how different parts of the bone relate to each other, even if they are far apart in the image.
• The Decoder (The Artist): Now, the AI starts rebuilding the image, piece by piece, from the bottom up.
• The Skip Connections (The Direct Line): This is the secret sauce. Usually, when an AI breaks an image down, it forgets the tiny details. HARU-Net builds "direct phone lines" (skip connections) from the Detective phase to the Artist phase.
  • The Hybrid Attention Block (HAB): Sitting on these phone lines are special filters. Imagine a smart spotlight. When the Artist is painting, the spotlight shines only on the important parts (the sharp edges of the bone) and ignores the irrelevant background (the air or noise). It tells the AI: "Hey, focus on this edge! Don't blur it!"

3. Why It's Better Than the Rest

The researchers compared HARU-Net to other top-tier AI models (like SwinIR and Uformer).

• The Competitors: These models are like heavy-duty supercomputers. They are incredibly powerful and accurate, but they are slow and require massive amounts of electricity (computing power) to run. They are like using a sledgehammer to crack a nut.
• HARU-Net: This model is like a precision Swiss Army knife. It combines the speed and efficiency of standard tools (Convolutional Neural Networks) with the smart, context-aware thinking of the heavy hitters (Transformers).
  • Result: It cleans the image just as well as the heavy hitters (actually, slightly better in some metrics) but does it much faster and with less computing power.

The Bottom Line

HARU-Net is a new, smart AI tool that can take grainy, low-radiation dental X-rays and make them look like high-definition, crystal-clear photos.

It does this by:

1. Learning from simulated "clean" jawbone scans.
2. Using a "hybrid" brain that looks at both tiny details and the big picture simultaneously.
3. Using "spotlights" to ensure it sharpens the bone edges without smearing them.

Why does this matter?
It means dentists and doctors can use lower radiation doses (safer for patients) and still get images sharp enough to see tiny fractures or plan complex surgeries. It's a faster, cheaper, and safer way to see inside the human body, bringing us one step closer to real-time, high-quality 3D imaging in the clinic.

Technical Summary

1. Problem Statement

Context: Cone-Beam Computed Tomography (CBCT) is a standard imaging modality in dentistry and maxillofacial/ENT diagnostics due to its high spatial resolution and lower radiation dose compared to medical CT.
Challenge: To minimize patient radiation exposure, low-dose acquisition protocols are used, which introduce strong, spatially varying noise (quantum and electronic noise). This noise degrades soft-tissue visibility and obscures fine anatomical structures (e.g., root canals, trabecular bone), leading to diagnostic uncertainty.
Limitations of Current Solutions:

• Classical Methods: Struggle to suppress noise without blurring critical edges.
• Deep Learning (DL) Methods: While promising, supervised DL approaches require paired high-dose (clean) and low-dose (noisy) data. Acquiring such pairs in clinical settings is ethically and practically impossible.
• Existing Datasets: Most available datasets are from radiotherapy (not dental/ENT) or rely on phantoms/cadavers that may not fully capture real anatomical heterogeneity. Self-supervised methods exist but often yield suboptimal performance.

2. Methodology

A. Data Preparation and Pre-processing

To overcome the lack of clinical paired data, the authors utilized a cadaver dataset of 21 human hemimandibles scanned using a high-resolution 3D Accuitomo 170 system (90 kV, 5 mA).

• Noise Simulation: Since the scans were high-dose (clean), the authors simulated low-dose noise by adding synthetic quantum noise (modeled as Gaussian based on Poisson statistics) and electronic noise to the clean images to create paired training data.
• Segmentation Pipeline: A specialized pre-processing pipeline was developed to isolate anatomical regions from background air:
  1. Manual Cropping: To exclude empty regions.
  2. K-Means Clustering: To separate foreground tissue from air based on intensity.
  3. Morphological Operations: Dilation to smooth boundaries and contour-based hole filling to remove internal air pockets.
  4. Dynamic Patching: Patches (256×256) were extracted exclusively from the segmented anatomical regions to ensure the model learns from relevant tissue rather than background noise.
• Dataset Size: 50,026 training pairs, 8,971 validation pairs, and 10,462 test pairs.

B. Proposed Architecture: HARU-Net

The authors propose HARU-Net, a Hybrid Attention Residual U-Net that integrates Transformer-based attention mechanisms into a Convolutional Neural Network (CNN) backbone. The architecture consists of:

1. Encoder: Four stages of residual convolutional blocks. Each block uses two 3×3 convolutions with LeakyReLU and a 1×1 projection for skip connections. Down-sampling is achieved via learnable strided convolutions (preserving more signal than pooling).
2. Hybrid Attention Transformer Block (HAB):
   • Location: Integrated into skip connections between the encoder and decoder.
   • Function: Combines Windowed Self-Attention (from Swin Transformer) to capture local spatial patterns and Channel Attention (Squeeze-and-Excitation) to re-weight feature channels based on global relevance.
   • Goal: To selectively emphasize salient anatomical features and refine features transmitted between encoder and decoder, preventing the loss of high-frequency details.
3. Residual Hybrid Attention Group (RHAG):
   • Location: Placed at the bottleneck (deepest layer).
   • Function: A series of six HABs in a residual configuration.
   • Goal: To strengthen global contextual modeling and long-range feature interactions where feature resolution is lowest.
4. Decoder: Mirrors the encoder using transposed convolutions for up-sampling. HABs are applied at each skip pathway to fuse and refine encoder features before reconstruction.

Training: The model was trained using Mean Squared Error (MSE) loss, the Adam optimizer, and an early-stopping strategy to prevent overfitting.

3. Key Contributions

1. Novel Architecture: Introduction of HARU-Net, which uniquely combines the local feature extraction efficiency of CNNs with the global context modeling of Transformers via HABs and RHAGs.
2. Edge-Preserving Denoising: The hybrid design specifically targets the preservation of high-frequency anatomical details (edges, textures) which are often lost in pure CNN or pure Transformer approaches.
3. Robust Data Strategy: Development of a rigorous pre-processing and noise-simulation pipeline using a high-resolution cadaver dataset to enable supervised training for dental CBCT, a domain previously hindered by data scarcity.
4. Efficiency: Demonstrating that a hybrid approach can outperform pure Transformer models (like SwinIR and Uformer) while maintaining significantly lower computational costs.

4. Results

The method was evaluated against a Residual U-Net (baseline), Uformer, SwinIR, and HAT (Hybrid Attention Transformer).

Quantitative Performance (Test Set):

• PSNR: HARU-Net achieved the highest score (37.52 dB), outperforming HAT (36.70 dB) and SwinIR (36.12 dB).
• SSIM: Achieved 0.9557, indicating superior structural similarity (slightly lower than HAT's 0.9569 but with better edge metrics).
• GMSD (Gradient Magnitude Similarity Deviation): Achieved the lowest score (0.1084), indicating the best preservation of edges and fine structural details.

Computational Efficiency:

• Inference Time: HARU-Net processed a full 3D scan (512³) in 1.985 minutes.
• Comparison: This is significantly faster than Uformer (4.30 min) and SwinIR (8.85 min).
• FLOPs: HARU-Net required 40.76 GMACs per patch, substantially lower than SwinIR (111.07) and HAT (349.36), though higher than the pure CNN baseline (ResU-Net at 6.90 GMACs).

Visual Assessment:
Qualitative analysis of sagittal, axial, and frontal views showed that HARU-Net produced the sharpest bone boundaries and most faithful trabecular patterns. Unlike other methods, it avoided over-smoothing and attention-induced artifacts.

5. Significance and Conclusion

• Clinical Impact: HARU-Net offers a practical solution for enhancing low-dose CBCT images, potentially reducing the need for repeat scans and improving diagnostic confidence for dental and ENT procedures.
• Architectural Insight: The study proves that integrating selective Transformer components (HAB/RHAG) into a CNN framework yields a "best of both worlds" scenario: the global context of Transformers with the computational efficiency and local feature stability of CNNs.
• Future Directions: While HARU-Net is faster than current SOTA Transformer models, it is not yet real-time. Future work will focus on model compression, 3D volumetric denoising, and domain adaptation to handle data from different CBCT vendors.

In summary, HARU-Net represents a significant advancement in medical image processing, successfully balancing high-fidelity denoising with computational feasibility for clinical deployment.