Initial convnet not learning anything about prostatex HOT 4 CLOSED

You can add the parameter stratify=labels to the train_test_split method. This way the train and validation set have the same unbalance.
Of the 334 lesions in the data set, 77 are clinically significant (~23%).

Are you using one-hot output?
i.e. 2 classes; not significant , significant?

Another question. Does the loss change?

I'm also trying to get a small network working and I think I am running into the same problem.

from prostatex.

I got the feeling that the input we are currently using may be the problem. When you look at the cutouts of the ADC scans of the lesions, I can't recognize anything that resembles a prostate or even a lesion.

Would that auto windowing help here?(not sure about the name) In making it more recognizable.
It can be that the problem (with the current input) is just too difficult for the network to start learning something.

from prostatex.

Yes right now it's using a one-hot representation. The loss changes a little bit. Sometimes a slight increase and sometimes a slight decrease. I'm also starting to suspect that using the lesions as is might not be the right approach. Windowing indeed might help the network focus better on what actually matters in the image.

from prostatex.

Input is not the problem.
For me the problem was using a softmax layer with only one output. (the sum of the output array always equals one...)

I also made some other changes that you can also try to get it working better. Use He initialization, use leaky ReLU activation function instead of ReLU to avoid dead neurons. And maybe initialize with slight positive biases, however I doubt that this really matters when you use leaky relu.

from prostatex.