First-hand information for everyone
Researchers showed that features learned by large pretrained models can be pulled from noisy cardiac MRI scans and improve reconstruction quality.
They built a system called FM‑Recon. It takes a frozen foundation model image encoder, tweaks its outputs with tiny adapters and channel attention, and merges the refined representations into each step of an unrolled reconstruction network using Cross‑FiLM modules. This fusion lets the model learn from both the foundation features and its own reconstruction features without adding much computation.
On the CMRxRecon2023 dataset, FM‑Recon beat a UNet‑MR baseline in PSNR, SSIM, DISTS, and HaarPSI for acceleration factors of 4×, 8×, and 24×. The improvement was largest at moderate acceleration (4×) and still present, though smaller, at higher rates. The gains varied with the chosen foundation model: DI‑NoV2 yielded the best results at 4× and 24×, MCP‑MedSAM excelled at 8×, and LiteMedSAM produced only modest benefits.
The study notes that when sampling is very aggressive, the benefit from foundation model features wanes, suggesting the feature refinement must be tuned to the undersampling level. It also highlights that the choice of foundation model matters for performance.
These findings demonstrate that pretrained model representations can be harnessed to speed up cardiac MRI recovery while improving image fidelity, especially when the undersampling is moderate.