A modular deep learning framework for accelerated MRI reconstruction that combines data-driven learning with physics-guided constraints. The dual-branch architecture employs a learned gating network to route k-space information adaptively, eliminating the need for anatomy labels during inference while maintaining high reconstruction fidelity across diverse anatomical regions.
The dual-branch architecture achieves marked improvements over single-branch baselines:
- Structural similarity improvement: +1.78% SSIM
- Temporal complexity: 112 milliseconds per inference
- Computational cost: 302 GigaFLOPs
Hardware & Evaluation Environment
- GPU: NVIDIA GeForce RTX 4060
- Precision: FP16 (Mixed Precision)
- Input Resolution: 320 × 320 (Complex domain)
- Inference Mode: 2D slice-by-slice
.
├── configs/ Configuration files for training and evaluation hyperparameters
├── notebooks/ Jupyter notebooks for interactive analysis and prototyping
├── results/ Experimental outputs, visualizations, and quantitative results
└── src/ Source code organized by functional module
| Name | GitHub | |
|---|---|---|
| Kunal | @Kunal-Somani | kunal-somani-227373344 |
| Vansh Chambyal | @Vansh-Chambyal | vansh-chambyal-6a6b62315 |