ResNet18 performance: Impact of network depth and image resolution on image classification

Liu, Haixia; Bull, Larry; Brailsford, Tim

Abstract

This study explores aspects of using ResNet18 for the classification of biomedical images, using the 2D medical images from the MedMNIST library that is widely used for benchmarking. The number of layers within the ResNet18 was studied. The typically used depth (block 4) was generally found to be robust across a variety of datasets at the recommended image resolution (224x224). We found that decreasing the resolution while maintaining block 4 depth can significantly improve performance in some, but not all, cases. The effects of varying both depth and resolution simultaneously were evaluated, and we found a non-linear relationship between depth/resolution and performance. The context of the images seems to be important, with the best performing combinations of network depth and image resolutions varying. We examined the feature maps, and found them to be very variable for the best performing models.

• Presentation Conference Type: Conference Paper (published)
• Conference Name: International Conference on Advances in Artificial Intelligence
• Start Date: Oct 17, 2024
• End Date: Sep 19, 2024
• Acceptance Date: Sep 5, 2024
• Deposit Date: Sep 5, 2024
• Publisher: Association for Computing Machinery (ACM)
• Peer Reviewed: Peer Reviewed
• ISBN: 9798400718014
• Keywords: CCS Concepts, Computing methodologies, Machine learning, Machine learning approaches, Neural networks
• Public URL: https://uwe-repository.worktribe.com/output/12842105
• Publisher URL: https://www.acm.org/publications/proceedings