Memory-efficient optimization of implicit neural representations for CT reconstruction
A novel gradient approximation technique enables high-quality 3D medical imaging on standard GPUs.
Researchers Mahrokh Najaf and Gregory Ongie have published a breakthrough paper addressing a critical bottleneck in medical imaging AI. Their work tackles the prohibitive GPU memory requirements of using Implicit Neural Representations (INRs) for CT reconstruction. INRs offer a parameter-efficient, fully differentiable model for creating images from CT scan data, but standard training methods require evaluating millions of virtual ray projections, consuming massive GPU memory—especially problematic for 3D imaging where memory demands scale exponentially.
The team's solution introduces a novel stochastic gradient approximation method that decomposes the gradient calculation into manageable Jacobian-vector products. This approach allows users to precisely control the trade-off between memory usage and reconstruction accuracy through strategic subsampling. In experiments on synthetic 2D data, their method achieved comparable mean squared error and convergence behavior to standard INR training while using dramatically less memory. Most significantly, they demonstrated practical 3D cone beam CT reconstruction in sparse-view settings—a previously computationally prohibitive task. This advancement could democratize advanced medical imaging research and clinical applications by enabling high-quality 3D reconstructions on more accessible hardware.
- Novel stochastic gradient approximation reduces GPU memory by decomposing gradients into Jacobian-vector products
- Enables 3D cone beam CT reconstruction in sparse-view settings previously requiring specialized hardware
- Maintains reconstruction quality comparable to standard methods while offering memory/accuracy trade-off control
Why It Matters
Makes advanced 3D medical imaging research and applications feasible on standard GPUs, potentially accelerating diagnostic AI development.