Deep Learning Reconstruction (DLR)

DLR algorithms are trained to distinguish between anatomical structures and quantum noise in projection data. Unlike traditional iterative reconstruction (IR) which often trades noise for a 'plastic' or 'waxy' texture, DLR preserves high-contrast spatial resolution while aggressively suppressing low-contrast noise.

• Mechanism: Trained on paired low-dose (noisy) and high-dose (clean) images.
• Result: Maintains the noise power spectrum (NPS) of high-dose scans at low-dose radiation levels.
• Key Metric: Noise reduction ratios of 50-70% compared to Filtered Back Projection (FBP) without degrading the modulation transfer function (MTF).

DLR excels at resolving non-linear artifacts that confound analytic and iterative methods. The network learns the characteristic appearance of artifacts directly from data, enabling it to suppress them without requiring explicit physical modeling.

• Photon Starvation: Corrects severe streaking in low-dose scans or through broad patient anatomy.
• Metal Artifact Reduction (MAR): DLR-based MAR normalizes corrupted projection data in the sinogram domain before reconstruction, outperforming interpolation-based methods.
• Beam Hardening: Reduces cupping artifacts by learning the non-linear relationship between polychromatic attenuation and monochromatic correction.

DLR can recover spatial frequencies beyond the traditional Nyquist limit of the acquisition detector. By learning a prior on high-resolution anatomical textures, the network hallucinates plausible fine details lost to the system's modulation transfer function (MTF).

• Application: Visualizing trabecular bone microarchitecture or small vessel lumens from standard-resolution acquisitions.
• Technique: Often employs a Generative Adversarial Network (GAN) loss or perceptual loss to enforce realistic texture synthesis.
• Caution: Requires rigorous validation to ensure 'hallucinated' edges correspond to true anatomy and not false positives.

Unlike generic image denoisers, medical DLR networks incorporate anatomy-aware priors. The model is trained exclusively on medical images, learning the expected texture, continuity, and topology of human tissue.

• Task-Specific Training: A network trained for lung CT learns to preserve fissures and ground-glass opacities; a brain MRI network preserves white/gray matter contrast.
• Contrast Preservation: Maintains the quantitative accuracy of Hounsfield Units (HU) for diagnostic measurement.
• Generalization Risk: Performance degrades if applied to anatomy or pathology not represented in the training distribution.

While the training phase is computationally intensive, the inference (reconstruction) speed of DLR is often faster than advanced model-based iterative reconstruction (MBIR).

• Architecture: Uses feed-forward convolutional neural networks (CNNs) or Vision Transformers (ViTs) that process data in a single pass.
• Comparison: MBIR requires hundreds of forward- and back-projection iterations; DLR applies a fixed set of matrix multiplications.
• Hardware: Optimized to run on dedicated Neural Processing Units (NPUs) or high-throughput GPUs integrated directly into the scanner reconstruction pipeline.

Advanced DLR architectures operate directly on raw sinogram or k-space data, bypassing the intermediate image domain entirely. This allows the network to correct inconsistencies in the raw acquisition data that are invisible after initial reconstruction.

• Sinogram DLR: Identifies and corrects corrupted detector readings before back-projection.
• K-Space DLR: In MRI, networks interpolate under-sampled k-space data to enable accelerated parallel imaging.
• Advantage: Solves the 'missing cone' problem in limited-angle tomography by learning to fill gaps in the Radon transform space.

The classic analytic CT reconstruction algorithm that applies a high-pass filter to raw projection data before back-projecting it across the image matrix. FBP is computationally fast but highly susceptible to noise amplification and streak artifacts, especially at low radiation doses. DLR fundamentally outperforms FBP by learning to suppress noise in the projection or image domain without sacrificing spatial resolution.

A computationally intensive technique that repeatedly compares forward-projected model estimates with raw acquisition data to reduce noise and artifacts. Unlike FBP, IR incorporates statistical models of photon distribution and system optics. DLR represents the next evolutionary step, replacing hand-crafted regularizers with deep neural networks trained on paired low-quality and high-quality data to resolve fine structures beyond IR's mathematical limits.

A signal processing framework that enables accurate image reconstruction from significantly under-sampled data by exploiting sparsity in a known transform domain, such as wavelets or total variation. Compressed sensing is a key enabler for accelerated MRI. DLR extends this concept by learning the optimal sparse representation directly from data, often achieving higher acceleration factors with fewer artifacts.

A class of algorithms designed to mitigate the severe streaking and dark-band artifacts caused by metallic implants like hip prostheses or dental fillings in CT images. Conventional MAR interpolates or normalizes corrupted projection data. DLR-based MAR approaches use neural networks to directly inpaint missing projection data or correct the reconstructed volume, yielding more anatomically faithful results.

The frequency-domain representation of an MR image, storing spatial frequency information acquired directly by the scanner's receiver coils. The center of k-space encodes image contrast, while the periphery encodes fine detail. DLR for MRI often operates directly on under-sampled k-space data, learning to reconstruct missing frequency components before transforming the result into a high-fidelity anatomical image via the Fourier transform.

A preprocessing step, often using the N4 algorithm, that removes low-frequency intensity non-uniformity artifacts inherent to MRI. These artifacts are caused by magnetic field inhomogeneities and coil sensitivity variations. DLR can be trained to perform joint reconstruction and bias field correction, learning to output a clean, uniform image directly from raw, distorted k-space data without a separate preprocessing pipeline.