Deep learning (DL) has shown significant potential in medical imaging, particularly in reconstructing high-quality CT and MRI images from sparse data. However, recent studies have highlighted critical instabilities in these models, including adversarial sensitivity and the loss of small structural features. This paper introduces the ACID (Analytic, Compressive, Iterative Deep) framework, designed to stabilize deep learning for tomographic reconstruction by embedding physical laws into the network's architecture.
The Challenge of AI Instability in Radiology
While DL models can reduce radiation doses and scan times, they often lack "kernel awareness"—an inherent understanding of the underlying physics, such as the Radon transform for CT or the Fourier transform for MRI. This leads to three primary issues:
- Adversarial Sensitivity: Minor input noise can result in significant output artifacts.
- Feature Loss: Critical details, like small tumors, may be omitted.
- The "More is Less" Paradox: Performance sometimes degrades as more input data is provided.
The ACID Framework
The ACID framework addresses these issues through a synergistic, iterative loop that combines traditional mathematical rigor with neural network performance.
The framework consists of four key components:
- Analytic Mapping: Uses known physical transforms to calculate the "residual" between current estimates and raw measurements.
- Deep Imaging: Employs pre-trained neural networks (e.g., Ell-50 or DAGAN) to refine image details based on large-scale data patterns.
- Compressed Sensing (CS): Applies Total Variation (TV) regularization as a physical "sanity check" to enforce sparsity.
- Iterative Feedback: Continually feeds errors back into the system, allowing the model to correct previous omissions.
The Superiority Principle
A core philosophy of this work is the Superiority Principle, which posits that the optimal solution lies in the integration of "classic" algorithms and "modern" AI.
By wrapping a neural network inside a physics-based iterative loop, the system remains grounded in physical reality. Unlike pure DL models, ACID's performance improves monotonically with additional data, resolving the "More is Less" paradox.
Experimental Validation and Results
The ACID framework was validated using simulated CT and MRI datasets designed to trigger known instabilities. Test images included phantoms with embedded "bird" logos and text ("CAN U SEE IT") to test feature recovery. Key results include:
- Robustness: ACID remained stable under adversarial attacks that caused standard DL models to produce artifacts.
- Feature Recovery: The system successfully reconstructed small text and logos that were erased by non-iterative networks.
- Quantitative Gains: ACID reduced Root Mean Square Error (RMSE) by 50-66% compared to standard CS or pure DL methods, while maintaining high PSNR and SSIM scores.
- Efficiency: Despite its iterative nature, the system is computationally efficient, with 50 CT iterations completing in under 35 seconds on a standard GPU.
Conclusion
The ACID framework demonstrates that embedding physical constraints into deep learning architectures is essential for reliable medical imaging. Future work will focus on automating hyperparameter tuning and exploring more advanced "deep priors" to further enhance reconstruction quality.
Summary updated