Hello,
Could you please clarify how the metrics are now being calculated with the masking?
As we understand a background mask (i.e., air) is generated per subject that is applied jointly to the enhanced prediction and high-field reference.
After this masking operation the metrics are calculated.
We are having difficulty in understanding the differences in the masked vs the unmasked scores.
Why does the PSNR increase but the MAE increase too?
Similarly, why does the SSIM decrease (when we'd expect it to increase due to more voxels being assigned an equal zero due to masking)?
Thank you!
Created by James Grover jgro4702525 Hi James,
Here are some clarifications to your questions.
**How metrics are calculated?**
To ensure fair and clinically relevant evaluation in this challenge, we compute metrics that focus only on the head region (brain including skull) , rather than the entire volume. We completely exclude the background in metric calculation and will release the script after the Docker evaluation.
**Why PSNR is higher than the conventional value ?**
Background regions typically contain low or zero intensity in HF and ULF images. So, any mismatch or noise in these low-intensity regions can significantly penalise PSNR in the conventional setting.
By excluding the background using the head mask, we remove irrelevant error contributions, highlighting true enhancement quality in the head region. Also, applying the head mask can increase MAE (due to higher absolute differences in structurally complex regions), while increasing PSNR (due to a reduction in small background errors that minimally affect MSE but penalize it numerically).
**Why SSIM Decreases?**
Including the background in SSIM computation increases the score because background voxels, which are mostly uniform, produce near-perfect SSIM values.
When the head mask excludes the background, SSIM decreases, because it now considers actual structural differences in the head area only, which reflects a more accurate assessment of image structural quality.
Best regards,
ULF-Enc 2025 Organizers