Issues processing 'large' images on a server

Recently we’ve switched to image analysis of whole slide, stitched images using Qupath (currently still on m8). Our images range from 4GB to 20GB. Our setup to analyze these large images is as follows: images placed on a ‘network’ hard drive are analyzed by running Qupath on a server with 128GB of RAM and 8-2.4GHZ Xeon CPUs. Admittedly, the images are not stored locally to where Qupath is running.

We find two major issues:

  1. In images ~>10GB, cell segmentation does not occur. Qupath gets hung up and no tiling occurs.
  2. In most images, that have successfully cell segmented, Qupath will crash after poking around a bit at the cell measurements (for instance subcellular detections).

So my questions are:

  1. Any recommendations for a ‘better’ setup to handle these larger images? Our preliminary testing does show that a computer with tons of RAM and locally stored images does seem to work ‘better’. However, multiple users being able to easily access a server seemed appealing to us.
  2. Has anyone else found success with their own setup for handling large images?


Hi @Colt_Egelston, I don’t have any experience of running QuPath at this scale myself. I’d say a lot depends upon what precisely you need to do with your analysis.

Regarding the major issues, I’d really need to see any log output (e.g. via View → Show log) - but only if it persists with the very latest milestone version (many things have been fixed along the way).

Personally, I’m reluctant to advise ‘more hardware’ as a solution to this kind of issue. You’re already working with much better specifications than I have regular access to :slight_smile:

I don’t generally recommend the subcellular detections much, partly because it is easy to generate an astronomical amount of data with them… I realise there are applications for which QuPath doesn’t currently have any suitable alternative command, but until anyone finds time to develop something better I’m afraid it’s a matter of living within its constraints (e.g. by analyzing smaller regions).

This isn’t related to basic Cell detection not running, but back in the day, my first big project involved FISH analysis of whole slide images, and the data files became too large quite rapidly. I ended up manually segmenting the tissue into parts. These days, it might be easier to split your tissue detection into 4 very large tiles, and run one tile per project.

With the easier scripting these days, that shouldn’t be too bad.

Might even be able to export all of the cells (after classification) separately from the subcellular detections, and recombine them all into one project. Depends on data size. Removing measurements used to help as well (Min and Max DAPI values in cells? shrug, toss it).