Training a detection classifier (normal vs. tumour)

Hi all,

I’m looking for some help/advice with qupath. I’m trying to train a classifier to do 2 things on a TMA using qupath 0.2.0-m7 on colon/ colon cancer (IHC-DAB for a protein expressed by lymphocytes and many epithelial nuclei).

  1. positive vs. negative cell detection (successful!)
  2. stroma vs. lymphoid aggregates vs. tumour vs. normal epithelium (unsuccessful!)

I have been successful with (1) however I’m struggling with (2).

My method so far follows: stain detection optimisation > TMA dearray > positive cell detection > add smooth features (FWHM 25um) > create detection classifier (using both un-smoothed and smoothed features and ‘intensity feature = nucleus:DAB OD mean’).

The detection classifier works for some TMA cores and is reasonable at differentiating stroma vs. lymphoid aggregates. However, I’m really struggling to get it to differentiate cancer vs. normal epithelium. I have trained it a lot (80-90 examples of tumour and normal epithelium) with limited success. Are my samples just too heterogenous? (I’m not doubting my own histopathology skills to differentiate normal from tumour!)

Anyone hints/tips/advice would be very helpful!

Thanks in advance

Hard to say without pictures showing what is working, what isn’t working, and what your base tissue looks like.
That said, it looks like you have discovered many of the feature options, but there are more. In case you haven’t looked into them: