I first noticed this issue in run_qualitative_analysis.py where it would error in main at line 315
concepts = [cg.ndata['concepts'].cpu().detach().numpy() for cg in cell_graphs]
because a cell graph wouldn't have a concepts data field.
I traced this back to graph creation and generate_cell_graphs.py, where the concepts not being created error was being silenced due to a try-except statement.
# e. build a kNN graph
try:
graph = knn_graph_builder.process(nuclei_map, features)
except:
print('Warning: {} failed during kNN graph building.'.format(image_path))
image_ids_failing.append(image_path)
pass
When I dug into this, I noticed that the nuclei_map, according to the second output of nuclei_detector.process(image), there are far fewer nuclei detected than there should be! I'm talking only 4 nuclei for BRACS_1897_DCIS_4.png!
Speaking of try-excepts silencing critical errors, when I isolated graph creation to just one graph, I noticed that the knn graph builder was failing as well due to the same only-4-nuclei-detected, but when creating all the graphs in series that error wasn't noticeable because the code was just using the knn graph from the prior ROI in the list instead of throwing a visible error. But that's a separate issue.
I first noticed this issue in
run_qualitative_analysis.pywhere it would error in main at line 315because a cell graph wouldn't have a concepts data field.
I traced this back to graph creation and
generate_cell_graphs.py, where the concepts not being created error was being silenced due to a try-except statement.When I dug into this, I noticed that the nuclei_map, according to the second output of
nuclei_detector.process(image), there are far fewer nuclei detected than there should be! I'm talking only 4 nuclei forBRACS_1897_DCIS_4.png!Speaking of try-excepts silencing critical errors, when I isolated graph creation to just one graph, I noticed that the knn graph builder was failing as well due to the same only-4-nuclei-detected, but when creating all the graphs in series that error wasn't noticeable because the code was just using the knn graph from the prior ROI in the list instead of throwing a visible error. But that's a separate issue.