output images larger the 512 x 512 come out of pixel clustering unrecognizable to original tiff image...

[jasonmadore]

Hi,
I have been using Docker images of the Mesmer segmentation and pixel clustering notebooks to process multiplex fluorescent images generated by Codex/Phenocycler. Prior to this, the images were processed and quality controlled using QuPath. These images are in tiff format and contain up to 27 channels, however, I am currently running as few as 10 channels through the Ark-analysis pipeline until I can optimize the process.

The 'segment_image_data.ipynb' notebook works well with both the example data set (512x512 pixels) and our data sets at 512x512 and 2048x2048 pixels. However, when I run the 'Pixie_cluster_pixels.ipynb' notebook with our data at 2048x2048 pixels, the resulting output is an unrecognizable cluster in comparison to the original TIFF image (see example image below). The pixel clustering works properly on the same TIFF image if it is cropped down to 512x512 pixels (see example below).

I have tried running each image multiple times by switching different channels from the TIFF file, but the resulting images are consistent overall. The notebooks are run with standard settings, without the optional code cells for blurring and filtering of cell compartments.

Any advice on how to proceed is much appreciated.

example images of the issue;

Segmenation followed by 'skewed' pixel clustering at 2048x2048 image

Segmentation followed by pixel clustering of 512x512 image works well.

[jasonmadore]

solved issue was with metadata for original tiff where pixel dimensions and resolution were mis-matched.