Abstract
Biomedical image reconstruction has undergone a significant transformation over the last five decades, moving from analytical mathematical frameworks to data-driven deep learning paradigms. This survey examines the evolution of these techniques, focusing on the transition from classical methods like Filtered Back Projection to modern architectures such as U-Nets and Transformers. By reviewing five years of research from the Machine Learning for Medical Image Reconstruction (MLMIR) workshop, we identify key trends in architecture, dataset utilization, and the ongoing challenges of computational complexity and model reliability in clinical settings.
1. Introduction
In medical imaging modalities such as Computed Tomography (CT) and Magnetic Resonance Imaging (MRI), data is not captured as a direct image but as raw signals representing projections or frequency-domain samples. Reconstructing the underlying anatomy from these signals constitutes an inverse problem.
These problems are often ill-posed, where incomplete or noisy data can lead to multiple possible solutions. Historically, this led to trade-offs between scan time and image quality. The advent of deep learning has provided new tools to regularize these problems using learned priors from large datasets.
2. Taxonomy of Reconstruction Methods
Reconstruction techniques can be categorized into three primary eras:
2.1 Classical Analytical Methods
These methods utilize direct mathematical transforms. A primary example is Filtered Back Projection (FBP), which relies on the Central Slice Theorem. While computationally efficient, these methods are sensitive to noise and require dense sampling patterns.
2.2 Iterative Methods
Iterative techniques, such as the Algebraic Reconstruction Technique (ART), frame reconstruction as an optimization problem. By iteratively minimizing the difference between the measured signal and the signal predicted by a forward model, these methods achieve higher quality in low-dose or sparse-view scenarios, albeit at a higher computational cost.
2.3 Deep Learning Paradigms
Modern reconstruction leverages neural networks to learn the mapping from signal space to image space.
- CNNs and U-Nets: The U-Net has become the standard architecture due to its ability to capture multi-scale features through an encoder-decoder structure with skip connections.
- Transformers and GANs: Recent shifts toward Transformers allow for modeling long-range dependencies, addressing global artifacts. GANs are employed to enhance perceptual quality and texture.
3. Data and Benchmarking
The proliferation of open-access datasets has been a catalyst for progress. Key benchmarks include:
- fastMRI: A large-scale dataset for accelerating MRI acquisition.
- BraTS: Focused on brain tumor segmentation in multimodal MRI.
- TCIA: Providing extensive CT and PET/CT data for oncology research.
4. Current Trends and Future Outlook
Analysis of recent literature indicates a move toward hybrid models that combine the local feature extraction of CNNs with the global context of Transformers. However, several challenges persist:
- Computational Complexity: The high memory requirements of 3D/4D Transformers limit their use in high-resolution volumetric imaging.
- Model Hallucinations: Ensuring that deep learning models do not introduce non-existent anatomical features is critical for clinical adoption.
- Physics-Informed Learning: There is a growing trend toward "unrolled" networks that incorporate the physical forward model into the neural network architecture to improve robustness.
5. Conclusion
The field of biomedical image reconstruction is currently dominated by deep learning approaches that offer superior performance over classical methods. Future research must address the trade-offs between performance and efficiency while ensuring the physical fidelity of reconstructed images for reliable clinical diagnosis.
Summary updated