Related Experiment Video
Updated: Oct 12, 2025
08:05
Lung CT Segmentation to Identify Consolidations and Ground Glass Areas for Quantitative Assesment of SARS-CoV Pneumonia
Published on: December 19, 2020
14.4K
Factors determining generalization in deep learning models for scoring COVID-CT images.
Michael James Horry1, Subrata Chakraborty1, Biswajeet Pradhan1,2,3
1Center for Advanced Modelling and Geospatial Information Systems (CAMGIS), Faculty of Engineering and Information Technology, University of Technology Sydney, Australia.
Mathematical Biosciences and Engineering : MBE
|November 24, 2021
Summary
Deep learning models for COVID-19 diagnosis show promise in generalizing to new datasets, achieving up to 86% F1 score. Key factors for successful generalization include uniform image acquisition and diverse CT slice positions.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- The COVID-19 pandemic spurred extensive development of AI models for medical image analysis.
- Many deep learning models for COVID-19 diagnosis lack proven generalization to diverse datasets, limiting clinical utility.
- Investigating model generalizability is crucial for translating AI tools into clinical practice.
Purpose of the Study:
- To assess the generalizability of deep learning models for COVID-19 diagnosis using cross-dataset validation.
- To evaluate the predictive performance of these models for COVID-19 severity on an independent dataset.
- To identify factors influencing the generalization capabilities of deep learning models in medical imaging.
Main Methods:
- Utilized publicly available COVID-19 Computed Tomography (CT) datasets for cross-dataset validation.
- Employed image preprocessing techniques including histogram equalization and contrast limited adaptive histogram equalization.
- Incorporated a learning Gabor filter to enhance feature extraction and model performance.
- Assessed model predictive accuracy for COVID-19 severity using an expertly stratified independent dataset.
Main Results:
- Deep learning models demonstrated successful generalization to external datasets under specific conditions, achieving F1 scores up to 86%.
- The top-performing model exhibited predictive accuracy ranging from 75% to 96% for lung involvement scoring on an independent dataset.
- Identified uniform training image acquisition and diverse CT slice positions as critical factors for enhancing model generalization.
Conclusions:
- Deep learning models can achieve significant generalizability for COVID-19 diagnosis when trained and validated appropriately.
- Image acquisition consistency and data diversity are paramount for developing robust AI diagnostic tools.
- Further research into factors influencing generalization can accelerate the clinical adoption of AI in medical imaging.
