Handling missing data in model input

[kurtulusbulus]

Hello,

Thank you for efforts in developing this excellent project and for making the code publicly available.

I’m currently applying your model to our own dataset, which involves recordings from a range of voltage indicators captured using a custom-built (non-commercial) image sensor. Due to the nature of the hardware, we perform substantial pre-processing before analysis. One issue we face is the presence of hot pixels (i.e. pixels that are always active regardless of the signal). We typically handle these by detecting them and replacing their values with either 0 or NaN, depending on the downstream task.

However, I couldn’t find any details, either in the paper, repository, or codebase (apologies if I’ve missed something), about how the model handles missing data or masked values. Could you advise on how best to integrate such pre-processed data, particularly in cases where certain pixels are intentionally removed or set to NaN?

Many thanks in advance for your time and support.

Best regards,
Kurtulus

[EOMMINHO]

Hi Kurtulus,

Thank you for applying our method to your dataset.

To get straight to the point: we have not explicitly handled hot pixels using our method before. However, if you replace the hot pixel values with 0, the code should likely still run as expected—assuming the number of hot pixels is not excessively high. Since we haven’t tested our algorithm on data with a significant number of hot pixels, we cannot say for certain where the exact limit lies.

That said, we do not recommend replacing hot pixels with NaN, as this can lead to unexpected issues during the conversion process to float32.

From a personal perspective, for more stable neural network training, I would suggest replacing hot pixel values with the average of all pixel values in your dataset rather than using 0.

Best,
Minho