HUMUS-Net Evaluation on fastMRI Brain Dataset

This repository contains an independent evaluation of the HUMUS-Net model applied to the fastMRI Brain dataset.

---

Summary

HUMUS-Net is a state-of-the-art hybrid unrolled multi-scale network architecture originally designed for accelerated MRI reconstruction of knee data.
This evaluation adapts and tests HUMUS-Net on the fastMRI brain dataset, using updated evaluation script.

---

Evaluation Contents

• evaluation/eval.py: Updated evaluation script customized for fastMRI brain data.
• evaluation/figures: A few visualization of a reconstruction examples.

---

Evaluation results

Evaluation on fastMRI Brain dataset (8x acceleration, 4% center k-space) results:

Method	SSIM	NMSE	PSNR
HUMUS-Net	0.8871	0.0265	31.9

• Cross-domain evaluation on fastMRI datasets shows HUMUS-Net achieves reasonable generalization from knee to brain.

Slice 08 reconstruction and error map:

---

How to Reproduce

1. Clone this forked repository.
2. Follow the original HUMUS-Net installation and setup instructions here.
3. Download and prepare the fastMRI brain dataset from fastMRI.
4. Run the evaluation using the updated evaluation/eval.py script:

---

Acknowledgments and references

• HUMUS-Net repository
• HUMUS-Net: Hybrid Unrolled Multi-scale Network Architecture for Accelerated MRI Reconstruction: Fabian, Z., Tinaz, B. and Soltanolkotabi, M. (2022) ‘HUMUS-Net: Hybrid unrolled multi-scale network architecture for accelerated MRI reconstruction’. arXiv. Available at: https://doi.org/10.48550/ARXIV.2203.08213.
• fastMRI repository
• fastMRI: Zbontar et al., fastMRI: An Open Dataset and Benchmarks for Accelerated MRI, https://arxiv.org/abs/1811.08839