0
$\begingroup$

I am building the neural network for image analysis to do Chest Xray classification (Abnormal/Pass).

The classification accuracy for abnormal Xray is low, I guess it is due to the lack of abnormal samples to train.

Attached image is the schema I built. Is there any ways to improve the classification accuracy for abnormal Xray?

enter image description here

$\endgroup$
  • $\begingroup$ Is it possible to increase the number of abnormal images by duplication and minor transmutations? e.g. flipping the images across the horizontal or vertical axes $\endgroup$ – StevenTheDataGuy Jul 17 at 13:59
0
$\begingroup$

When the data is highly imbalanced, accuracy could be a bad measure of model performance. Instead consider precision and recall values, which separately takes into account the no of positive and negative samples. One idea to improve the performance is to use class weights (generally inversely proportional to the no of samples in that class) while calculating loss metric. This makes the model understand that class with less samples should be given more priority while training.

$\endgroup$
  • $\begingroup$ I would also advocate for using ROC-AUC as it is non local in the space of threshold. $\endgroup$ – Learning is a mess Jul 17 at 14:03
0
$\begingroup$

For anomaly detection problems, i'd recommend treating it some other way by training only on the normal samples and in prediction mode , classify samples that are normal as normal and flag everything else as abnormal. This approach is usually followed when you don't have enough "abnormal" samples in any anomaly detection problem to train on. Plus , no matter how much you train your model on anomalies, your model could face an unseen anomaly and pass it as normal and that could be fatal ( in healthcare especially ). You can use your abnormal samples as test samples in a hold-out set as a final evaluation step for your model.

Here's Andrew Ng's take on anomaly detection systems

$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.