I created a multi-label classification CNN to classify chest X-ray images into zero or more possible lung diseases. I've been doing some configuration tests on it and analyzing its results and I'm having a hard time understanding some things about it.

First of all, these are the graphs that I got for different configurations:

Results of CNN with different configurations

Note 1: I've only changed the dataset size and the number of color channels in each configuration Note 2: In case you're wondering why I tested the network with both 1 and 3 color channels, it's because the images are technically grayscale, but I am using the AlexNet architecture, which was made to take as input 224 x 224 images with 3 channels, so I wanted to see if the network somehow performed better with 3 channels instead of just the one

These are the things about it I don't understand:

  1. Why does the sensitivity and specificity of the network vary so much between different epochs?
  2. Is it normal for the validation loss of the network barely ever change as the number of epochs increase?
  3. Looking at the results I got, it looks like 2 epochs is where there tends to be the best results. Does that make sense? I've heard of people training their networks with dozens of epochs sometimes.
  4. Why is it that, many times, when the sensitivity of the network increases between epochs, the specificity tends to decrease, and vice-versa?

Sorry if some of these questions are dumb, I'm still a newbie at this. Also, my total dataset is drastically larger than what I present in these results (~110,000 images). I just haven't done tests with more images due to the time the network takes to train.

Network Architecture:

  • Base Architecture: AlexNet
  • Loss Function: Sigmoid Cross-Entropy Loss
  • Optimizer: Adam Optimization Algorithm with learning rate of 0.001

EDIT: I forgot to mention that the number of diseases to predict is 15, and that the network sees 0's much more than 1's due to the imbalance of classes. I've considered changing the loss function to a weighted version of sigmoid cross-entropy because of that, but I'm not sure if that would help the network much.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.