Abstract
Generating synthetic Computed Tomography (sCT) images from Cone Beam Computed Tomography (CBCT) is a critical task for improving the accuracy of radiotherapy planning and clinical diagnostics. Traditional methods using Convolutional Neural Networks (CNNs) or Transformers often struggle to balance local texture preservation with global anatomical context. This paper introduces the Global-Local Feature and Contrast (GLFC) learning framework, which utilizes a Mamba-Enhanced UNet (MEUNet) and a Multiple Contrast Loss (MCL) function. By integrating Mamba blocks into the skip connections of a high-resolution UNet, the model captures long-range dependencies with linear complexity. Furthermore, the MCL supervises the model across multiple intensity windows to ensure high-fidelity reconstruction of both soft tissues and bone. Experimental results on the SynthRAD2023 dataset show that GLFC significantly improves image quality, raising the Structural Similarity Index (SSIM) from 77.91% to 91.50% while maintaining a compact parameter footprint of just 5.57MB.
Introduction
In modern radiotherapy, Computed Tomography (CT) is the gold standard for dose calculation due to its accurate Hounsfield Unit (HU) values, which represent tissue density. However, clinics frequently use Cone Beam CT (CBCT) for daily patient positioning because of its lower cost and radiation dose. The primary drawback of CBCT is its susceptibility to noise, scatter artifacts, and inaccurate HU values, which limit its utility for precise treatment planning.
Deep learning models for CBCT-to-CT translation typically face a trade-off. CNNs excel at capturing local details but lack the receptive field to understand global anatomical structures. Conversely, Transformers provide global context but are computationally intensive and can blur fine-grained details. The GLFC framework addresses these challenges by combining the efficiency of State Space Models (Mamba) with a multi-window loss strategy designed to mimic clinical radiological assessment.
Methodology
1. Mamba-Enhanced UNet (MEUNet)
The core of the GLFC framework is the MEUNet, a high-resolution architecture designed to minimize information loss during down-sampling. Unlike standard UNets that rely on deep pooling layers, MEUNet uses only two down-sampling stages to preserve spatial resolution. To compensate for the reduced depth, the authors integrate Visual State Space (VSS) blocks, or Mamba blocks, into the skip connections.
In this configuration, the Mamba blocks act as a bridge between the encoder and decoder. They process the low-level features from the encoder to capture long-range, global dependencies before they are concatenated with the up-sampled features in the decoder. This allows the model to maintain local precision through the convolutional backbone while gaining the "big picture" awareness typically reserved for Transformers, but with significantly lower computational overhead.
2. Multiple Contrast Loss (MCL)
Standard loss functions often fail in medical imaging because they treat the entire intensity range equally. In a CT scan, the high intensity of bone can dominate the loss calculation, leading the model to ignore subtle contrast variations in soft tissues like the brain or muscles.
The proposed Multiple Contrast Loss (MCL) addresses this by calculating synthetic loss across three distinct intensity windows:
- Global Window: Covers the entire HU range to maintain overall structural integrity.
- Soft Tissue Window: Focuses on a narrow, low-intensity range to enhance contrast in brain matter and organs.
- Bone Window: Focuses on high-intensity ranges to sharpen the edges of the skull and skeletal structures.
By supervising the model in these specific ranges, GLFC ensures that the resulting sCT is optimized for both dense and soft anatomical features.
Experimental Setup and Results
Dataset
The framework was evaluated using the SynthRAD2023 Grand Challenge dataset, consisting of 180 patient cases with paired 3D head and neck CBCT and CT scans. The 3D volumes were sliced into 28,631 2D images (256×256 resolution) for training and validation.
Quantitative Performance
GLFC demonstrated superior performance across all major metrics compared to baseline CBCT and existing state-of-the-art models like SwinUnet and UNet++:
- SSIM: Improved from 77.91% (CBCT) to 91.50% (GLFC).
- PSNR: Increased from 20.09 to 29.84.
- Soft Tissue Fidelity: In the specific soft-tissue range, GLFC achieved an SSIM of 56.83%, outperforming UNet++ (51.79%).
Model Efficiency
One of the most significant advantages of the MEUNet is its efficiency. The model contains only 5.57 million parameters, making it approximately nine times smaller than UNet++ (47.1M parameters). This reduction in size makes the model highly suitable for deployment in clinical environments where hardware resources may be limited.
Discussion and Conclusion
The GLFC framework represents a significant step forward in synthetic CT generation. By strategically placing Mamba blocks in skip connections, the model successfully bridges the gap between local and global feature extraction without the quadratic complexity of Transformers. The introduction of Multiple Contrast Loss further ensures that the model remains sensitive to the subtle tissue gradients necessary for clinical accuracy.
Future work will focus on extending the MEUNet to a native 3D architecture to improve inter-slice consistency and validating the framework on other anatomical regions, such as the pelvis and abdomen. Ultimately, these high-fidelity synthetic CTs could enable more accurate "dose-of-the-day" calculations, improving the precision of adaptive radiotherapy.
Summary updated