|
Activity Number:
|
471
- Contemporary Statistical Methods for Imaging Data Analysis
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Thursday, August 6, 2020 : 10:00 AM to 2:00 PM
|
|
Sponsor:
|
Section on Statistics in Imaging
|
|
Abstract #309536
|
|
|
Title:
|
Pill Shape Classification using Imbalanced Data with Human-Machine Hybrid Explainable Model
|
|
Author(s):
|
William Lamberti*
|
|
Companies:
|
George Mason Univeristy
|
|
Keywords:
|
Pill Shape;
Classification;
Shape Metric;
Imbalanced Data;
Small Data;
human-machine hybrid
|
|
Abstract:
|
This paper presents a highly accurate interpretable solution for pill shape classification. We arrived at a human-machine hybrid approach that achieved an overall classification rate of 97.83% and a mean precision of 98.4%. The only misclassifications occurred between ovals and capsules. This corresponds to an average outperformance of 94% compared to the results of other approaches when using mean precision. Our final model used a decision tree where each node classified meta-classes, or groups of classes. Each node used support vector machines with a polynomial kernel. The tree was able to overcome imbalanced data between the classes by using meta-classes. Each node of the decision tree was limited to using only two variables. This made each node interpretable as the final decision boundaries can be plotted on a 2D scatterplot.
|
Authors who are presenting talks have a * after their name.
