Label co-occurrence warning

[Caleb Robinson]

This dataset is set up in a multi-label way, however most labels co-occur. The below table shows counts of unique label combinations over the 514 images in the AP and PA views. You can see, for example, that "Pneumonia" occurs in every row that has a label and is not labeled as "No Finding" (and furthermore, the second most common configuration is no label).

('COVID-19', 'Pneumonia', 'Viral') 320
() 37
('Pneumonia',) 31
('Fungal', 'Pneumocystis', 'Pneumonia') 22
('Pneumonia', 'SARS', 'Viral') 16
('No Finding',) 14
('Bacterial', 'Pneumonia', 'Streptococcus') 13
('ARDS', 'COVID-19', 'Pneumonia', 'Viral') 11
('MERS', 'Pneumonia', 'Viral') 8
('Lipoid', 'Pneumonia') 8
('Bacterial', 'Klebsiella', 'Pneumonia') 8
('Bacterial', 'Legionella', 'Pneumonia') 7
('Bacterial', 'Mycoplasma', 'Pneumonia') 5
('Pneumonia', 'Varicella', 'Viral') 5
('Influenza', 'Pneumonia', 'Viral') 3
('ARDS', 'Pneumonia') 3
('Bacterial', 'Pneumonia') 2
('Bacterial', 'Chlamydophila', 'Pneumonia') 1

The README shows aggregate label counts; maybe it is worth including a summary like above too in order to warn users who may want to use the excellent Dataset class for fine-tuning?

[Joseph Paul Cohen]

The labels are designed to be hierarchy. This is shown in the table 1 from the paper:

So they are supposed to co-occur because Pneumocystis is a Fungal Pneumonia.

the second most common configuration is no label

There shouldn't be any ones missing a finding. There are ones labelled as todo which means they are currently being studied.

maybe it is worth including a summary like above too

Maybe including something like table 1 will make it clear?