Keywords: Classification, Deep Learning, Medical imaging
TL;DR: Multiclass extension of the GEV activation to improve unbalanced multiclass classification,
Abstract: Unbalanced data poses a challenge when training machine learning algorithms; the algorithm often overfits on the dominant class and neglects the smaller classes. While methods such as oversampling aim to rebalance the data, this can lead to overfitting. When a certain class is underrepresented, either because it a rare disease or few images exist then methods are needed which can adequately account for this. The generalized extreme value (GEV) activation has recently been proposed as a solution to highly unbalanced data; however, the GEV activation is only available for binary classification. We extend this to the multiclass case with the multiclass GEV (mGEV) activation. We conduct experiments on X-ray images, with three classes, showing much-improved performance over the commonly used softmax activation. Code for the mGEV activation is available at [https://github.com/JTBridge/GEV].
Paper Type: methodological development
Primary Subject Area: Detection and Diagnosis
Secondary Subject Area: Learning with Noisy Labels and Limited Data
Paper Status: original work, not submitted yet
Source Code Url: https://github.com/JTBridge/GEV
Data Set Url: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/DBW86T
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
Source Latex: zip