Question Regarding Multi-Band EPI Reconstruction and Intensity Normalization

[mfaizan-ai]

Dear Author,

I hope this email finds you well.

My name is Faizan, and I am a PhD student at Trinity College Dublin. My research focuses on motion-robust EPI reconstruction. I recently came across your presentation on MR reconstruction, and I found it extremely insightful and helpful for my work.

I am currently working with Siemens raw data acquired from infant subjects. My first objective was to reconstruct the multi-band data into Siemens-like single-band images, which I have successfully achieved. However, I have encountered a few issues that I would greatly appreciate your guidance on.

First, compared to the Siemens-reconstructed images, my reconstructions appear less homogeneous in intensity and slightly less smooth. While the anatomical structures are consistent, there is noticeable intensity inhomogeneity (I will attach a comparison image for reference).

Second, there is a substantial difference in intensity scaling. My reconstructed images have values in the range of approximately 0–0.4, whereas the Siemens images range from 0–400. I would like to compare the two reconstructions using metrics such as SSIM, PSNR, RMSE, and NCC. What would be the appropriate way to bring both datasets to a comparable intensity scale before performing these evaluations?

In terms of my current reconstruction pipeline, I am performing phase correction and noise decorrelation. My advisor suggested applying prescan normalization for intensity correction. I attempted this using the body coil and head coil prescan information available in each .dat file; however, I suspect I may not be handling the geometry or interpolation correctly, as the results were not satisfactory.

As an alternative, I applied bias field correction using the SimpleITK Python module. This improved the visual similarity to Siemens images, but I am unsure whether this is the correct approach or whether it would meaningfully improve quantitative metrics.

Could you kindly advise on:

Additional preprocessing or post-processing steps that might be necessary in the reconstruction pipeline?

The correct approach to intensity normalization and prescan handling?

Best practices for fairly assessing reconstruction quality against vendor images?

Any open-source reconstruction frameworks or implementations that may be helpful in this context?

Your guidance would be greatly appreciated.

Thank you very much for your time, and I look forward to hearing from you.