Examination of the Double Descent Phenomenon in Medical Imaging AI Restricted; Files Only

Abstract

One of the reasons behind the success of neural networks in radiology and other fields (e.g.: natural language processing) is that the models learn complex features from the training data and generalize well to new data. The flexibility and success of the models is due to over-parameterization. Generally, in order to produce an over-parameterized model, researchers use a simple heuristic: the number of parameters in the model should be much greater than the number of training samples. Given the differences between natural and medical images, this heuristic when applied by researchers in the medical imaging community, might lead them to develop critically parameterized or under-parameterized models that lead to worse performance. As such, in this thesis, we aim to investigate whether a commonly used model (Densenet 121 model) in medical imaging research is under-parameterized or over-parameterized for a certain combination of factors. These factors include transfer learning, data set size, data complexity and model width. We restricted the task of the model to a simple binary classification of disease from a patient chest radiograph. We find that for certain training sizes, the model is in the critically parameterized regime and a tenfold increase in the sample size does not yield better performance. We also find that diseases that are more challenging to diagnose (such as COVID-19) typically shift the interpolation threshold to the right and cause the model to become over-parameterized. Given these results, we find that it is important for researchers in the medical imaging field to not use heuristics as the ones researchers in computer vision use to develop deep learning models for natural images.

Table of Contents

1. Introduction                                                                                                        1

     1.1 AI and its Impact on the Radiological Field....................................1

     1.2 The Schism .................................................................................1

     1.3 What is Double Descent?..............................................................2

     1.4 Double Descent in Practice...........................................................2   

     1.5 Radiology vs. Natural Image Classification ....................................4

     1.6 Aims of Thesis.............................................................................5   

     1.7 Related Works .............................................................................6

     1.8 Notes on the History of Double Descent ........................................7

2. Materials and Methods                                                                                        9

     2.1 Datasets ......................................................................................9

     2.2 Model Architecture .....................................................................10

3. Classifier Performance for COVID-19 and Pneumonia                                11

     3.1 Experimental Protocol ................................................................12

     3.2 AUCROC ....................................................................................13

4. Pretraining and Double Descent                                                                        15

    4.1 Experimental Protocol...................................................................15

     4.2 Results and Discussion..................................................................17

5. Impact of Model Width, Data Size on Double Descent                                  19

    5.1 Experimental Protocol....................................................................19

     5.2 Results and Discussion...................................................................21

6. Conclusion and Future Direction                                                                         24

About this Honors Thesis

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Emory College
Department	Physics
Degree	B.S.
Submission	Honors Thesis
Language	English
Research Field	Computer Science Information Science Physics, General
Stichwort	Artificial Intelligence (AI) Double Descent Medical Imaging
Committee Chair / Thesis Advisor	Justin Burton, Emory University
Committee Members	Tankut Can, Emory University Ilya Nemenman, Emory University

Zuletzt geändert

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	File download under embargo until 22 November 2025	2025-04-08 21:34:11 -0400	File download under embargo until 22 November 2025

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions