Related Experiment Video
Updated: Dec 12, 2025
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
9.8K
Comparing different deep learning architectures for classification of chest radiographs.
Keno K Bressem1, Lisa C Adams2, Christoph Erxleben3
1Charité Universitätsmedizin Berlin, Campus Benjamin Franklin, Hindenburgdamm 30, 12203, Berlin, Germany. keno-kyrill.bressem@charite.de.
Scientific Reports
|August 14, 2020
Summary
Shallow convolutional neural networks (CNNs) show comparable performance to deep networks for classifying chest radiographs. These simpler models offer faster training and efficient medical image analysis, even on limited hardware.
Area of Science:
- Radiology
- Medical Imaging
- Computer Vision
Background:
- Chest radiographs are common in radiology and a focus of computer vision research.
- Existing models often use deep neural networks trained on diverse, large datasets not optimized for medical images.
Purpose of the Study:
- To compare the classification performance of 16 different convolutional neural network (CNN) architectures on medical imaging datasets.
- To determine if simpler, shallower CNNs can achieve state-of-the-art results in chest radiograph classification.
Main Methods:
- Evaluated 16 CNN architectures on two public datasets: CheXpert and COVID-19 Image Data Collection.
- Assessed classification performance using Area Under the Receiver Operating Characteristics Curve (AUROC).
Main Results:
- AUROC values between 0.83-0.89 were achieved on the CheXpert dataset.
- Excellent detection of COVID-19 and non-COVID pneumonia with AUROC values of 0.983-0.998 on the COVID-19 dataset.
- Shallower networks demonstrated comparable performance to deeper networks with reduced training times.
Conclusions:
- Simpler, shallower CNNs are effective for chest radiograph classification.
- These models offer efficient medical image analysis, achieving high performance comparable to complex models, even with limited resources.
