Back to search

University of South Florida

Autoencoders as Classifiers Trained on Single Sources of Radiographic Images for Generalizability across Unseen Sources

Abstract

dc:description

Artificial neural networks trained to classify X-ray images according to disease state will learn to distinguish between technical or procedural variations in the images rather than features relevant to the disease. For example, the model might learn to recognize the machine that captured the image or to distinguish an image taken while the patient is lying supine or standing upright. The present work is aimed at using autoencoders as a generalizable classifier to detect COVID vs pneumonia (PNA) X-ray images. The first experiment was to find an architecture for a fully convolutional autoencoder (CAE) and for a convolutional autoencoder with a dense, fully connected middle layer (CDAE) that reconstructed the input images with as low an error as possible. First, several different CAEs were trained on a data set consisting of one source of COVID X-rays and one source of PNA X-rays. The best performing CAEs were then given a dense, fully connected inner layer and trained on the same data set. The CAE/CDAE pair with the least overall loss were chosen to be used in the following experiment. The second experiment was to use a pair of autoencoders with the same architecture as a classifier. One was trained on a single source of COVID images and the second on a single source of PNA images. The COVID and PNA images from sources other than the ones used for training comprised the test set. If the reconstruction loss for the COVID-trained autoencoder was less than that for the PNA-trained autoencoder, the image was classified as COVID and vice versa. Classifiers tested on images from the same source as the training set can achieve a perfect AUC of 1.0, but only have an AUC of 0.38 when tested on unseen sources [1]. In this research the best performing pair of CAEs achieved an overall accuracy of 62.5% with class-level F1-scores of 0.61 for COVID and 0.64 for PNA, and the best performing pair of CDAEs achieved an overall accuracy of 68.7% with class-level F1-scores 0.55 for COVID and 0.76 for PNA.

Degree

thesis:*
Name thesis:degree_name
MS in Computer Science (M.S.C.S.)
Level thesis:degree_level
Thesis
Discipline thesis:degree_discipline
Computer Science and Engineering
Grantor
University of South Florida
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Putney, Ryan M.
Contributors dc:contributor
  • Lawrence Hall, Ph.D.
  • Dmitry Goldgof, Ph.D.
  • John Templeton, Ph.D.

Subjects

dc:subject × 6

Identifiers

dc:identifier.*
Repository record dc:identifier
https://digitalcommons.usf.edu/etd/11391

Chain of custody

source
Harvested from
University of South Florida
Base URL
digitalcommons.usf.edu/do/oai/
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Putney, Ryan M.. Autoencoders as Classifiers Trained on Single Sources of Radiographic Images for Generalizability across Unseen Sources. Thesis thesis, University of South Florida, 2026. https://digitalcommons.usf.edu/etd/11391