All Times ET
Keywords: Shape Classification, Medical Pills, Machine Learning Applications
This paper presents a competitive solution for pill shape classification using support vector machines (SVM) models with interpretable features. Our first experiment shows the our SVM model outperforms convolutional neural network (CNN) approaches for pill shape classification. Our second experiment provides a human-in-the-loop approach that was able to achieve an overall classification rate of about 97.83% and a mean precision of 98.4%. The human-in-the-loop component only decided which variables to use for a given pair of meta-classes, or groups of classes. Our multinomial classification problem was converted into a series of binary classification problem through the use of meta-classes. Furthermore, by using variables which have a clear meaning, we created a model which is far more interpretable than its CNN counterparts. The code is available at https://github.com/billyl320/human_decision_tree_pills and https://github.com/billyl320/SPEI-Paper/tree/nml_nih.