I found it more interesting (and usefull for the task) that images are available. Otherwise it’s just another train-a-tree task. Features can be very affected by noise (as showed in the feature description) or it may be not completelly representative for the task
Totally agree with you - feature engineering should be left to the data scientists to experiment as they want.
It’s unfortunate that they don’t share the images.
Automatic feature extraction could be more performant than all those manual selected features combined.
Unfortunately, images can not be made available at this time because of data sharing restrictions
It’s unfortunate really. The framing of this problem practically assumes that the DL revolution didn’t happen.
Looking at it another way, it assumes that it has happened and that DL models are good enough to reliable extract features, thus allowing us to play with data that we would otherwise never have access to for legal reasons.
If the image -> diagnosis pipeline appeals, you cal always use a drawing library to reconstruct a plausible clock face