This repository presents the official implementation of FSSCAN (Frequency–Spatial Skip Connection Attention Network) for accelerated MRI reconstruction. FSSCAN is a complex-valued deep learning architecture designed to effectively exploit both frequency-domain and image-domain information for high-quality MRI recovery.
In addition to the proposed FSSCAN model, this repository includes implementations of several state-of-the-art (SOTA) MRI reconstruction methods, comprehensive ablation studies, and reusable complex-valued convolutional neural network modules.
- Accelerated MRI reconstruction using deep learning
- Comparison of proposed FSSCAN with SOTA methods
- Ablation study for architecture analysis
- Proper handling of complex-valued MRI data
- Ablation study/ – Ablation experiments for model analysis
- DCRCNN_code/ – DCR-CNN implementation
- DMSENet_code/ – DMSENet implementation
- Data/ – undersampling masks
- FSSCAN_Revision/ – FSSCAN experiments
- Modules-20260202.../Modules – Complex-valued neural network components
- RNLF_code/ – RNLF model implementation
- SOTA_paper_2_HFGN/ – HFGN implementation
- TEID_Code/ – TEID-Net implementation
- Unet_code/ – U-Net baseline implementation
Each model is organized in a separate folder with training and testing scripts.
The dataset subsets used for ablation studies (including training, and validation) can be downloaded from the following Google Drive link:
👉 train : https://drive.google.com/drive/folders/1uqKEVBWeeDOZv3QKsYBA_4GHt0uww7hK?usp=drive_link
👉 Val : https://drive.google.com/drive/folders/1syxbiZyVPZcFCcw8q15TMIBSoTnOGlsL?usp=drive_link
Checkpoints for different models are available on https://drive.google.com/drive/folders/1qenp9ijyVZNWlp3j-O0FigkZnclKoZEM?usp=drive_link