ResNet18 Performance: Impact of Network Depth and Image Resolution on Image Classification
Pages 351 - 356
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.
1 Introduction
The usual approach to image classification using ResNet18 [5] and ImageNet [2] involves resizing images to 224x224 before processing with a convolutional neural network (CNN). The prevailing view is that this is a good compromise between the amount of information in the image and the computational requirements of training [4]. Despite the widespread use of this size, there is little consensus that it always produces the best results and it is quite likely that the context of use is important. It is usually assumed that deeper networks perform better than shallow ones, and Shamir [11] has provided a theoretical justification as to why this is.
The debate about the relative effectiveness of deep versus shallow learning is ongoing, with various studies offering different perspectives. For example, Robles Herrera et al. [9] have suggested that deep learning is more effective in situations with natural symmetries, while shallow learning may be better in other cases. Malach and Shalev-Shwartz [6] have explored the relationship between the expressivity of deep networks and their training efficiency, finding that the efficacy of deep networks depends on whether the distribution can be well approximated by shallower networks. Mhaskar et al. [7] have provided a theoretical basis for the effectiveness of deep convolutional networks in approximating compositional functions with fewer parameters. Furukawa and Zhao [3] have focused on the interpretability of deep learning, finding that simpler and better rules can be extracted from higher layers of deep networks. Picon Ruiz et al. [8] have provided a broader perspective, explaining that deep learning often outperforms more traditional machine learning due to its ability to handle complex pattern recognition tasks. However, Cirrincione et al. [1] have found that for a gear fault diagnosis task, shallow networks generally outperform deep ones. Recent studies by Zhang et al. [13] have incorporated dendritic learning into convolutional neural networks, in an attempt to more closely mimic biological neurons’ information processing capabilities. They have used a shallow architecture of pyramidal neurons for image classification. Despite all of this, the relationships between image size, network depth and performance are still not well understood, especially in medical images, rather than the more commonly studied ImageNet images [2].
In this paper, the research questions that we have been addressing are:
•
For each MedMNIST dataset explored in this study, are there better, shallower depths rather than 4 blocks based on 224x224 resolution of the standard version of ResNet18?
•
For each MedMNIST dataset explored in this study, are there better image resolutions in comparison to 224x224 based on a depth of 4 blocks?
•
For each MedMNIST dataset explored in this study, are there better combinations of depths and resolution rather than using the 4 blocks and 224x224 resolution?
•
What are the feature map sizes of the final CNN layer in the best performing models? Do they maintain the same size, or do they vary? If they differ, what might be the underlying reasons?
2 Methodology
2.1 Datasets
Several 2D medical image datasets (MedMNIST) [12] were used in this study, which have been widely used for example in Zhang et al. [13]. Table 1 shows the details of these datasets, with sample size indicating the total number of samples in each. These datasets vary in size and number of classes, with up to 58,850 samples and 11 classes. Images were pre-processed utilizing several techniques including resizing, normalisation.
Image resizing: The 2D medical image data were obtained from the MedMNIST library through its API 1. The images used were already of the sizes that were used in the original study [12] (e.g.: 28x28, 64x64, 128x128 and 224x224).
Normalization: Input values were normalized using the transforms library2, with mean and standard deviation parameters set to 0.5. This process involves subtracting the mean and dividing by the standard deviation for each channel, which is one of the basic pixel value scaling techniques [10].
Training: We used stratify=y during data splitting to ensure that the class distribution is maintained in the training, validation and testing sets. Sixty percent of the data was used for training, twenty percent for validation and twenty percent for testing.
Table 1:
| Dataset name | Sample size | Image Type |
| BloodMNIST | 17,092 (8 classes) | color |
| BreastMNIST | 780 (Binary-Class) | gray |
| DermaMNIST | 10,015 (7 classes) | color |
| PneumoniaMNIST | 5,856 (Binary-Class) | gray |
| RetinaMNIST | 1,600 (5 classes) | color |
| OrganAMNIST | 58,850 (11 classes) | gray |
| OrganCMNIST | 23,660 (11 classes) | gray |
| OrganSMNIST | 25,221 (11 classes) | gray |
2.2 Feature Map Size
The fundamental unit utilised for building ResNet18 is called a “basic block”, which is a concept introduced in the original ResNet18 paper [5], and illustrated in Table 3.
Unlike many other studies that go deeper using compound scaling, we go shallower and down scale the image resolution, in an attempt to achieve similar results using less computational power. Table 2 shows the feature maps size per block per image resolution.
Table 2:
| Resolution | Block1 | Block2 | Block3 | Block4 |
| 28x28 | 7x7 | 4x4 | 2x2 | 1x1 |
| 64x64 | 16x16 | 8x8 | 4x4 | 2x2 |
| 128x128 | 32x32 | 16x16 | 8x8 | 4x4 |
| 224x224 | 56x56 | 28x28 | 14x14 | 7x7 |
Table 3:
| Layer | Description |
| conv1 | Convolutional layer with kernel size 3x3, padding of 1 |
| bn1 | Batch normalization layer |
| conv2 | Convolutional layer with kernel size 3x3, padding of 1 |
| bn2 | Batch normalization layer |
| relu | Rectified Linear Unit (ReLU) activation function |
| identity_downsample | Optional downsampling function |
2.3 Experiment Settings and Design
The experiments were implemented in Python using Pytorch3. During training, all the experiments were done under the same settings: the maximum number of epochs was set to 50; batchsize was 32; Stochastic Gradient Descent (SGD) optimizer with a learning rate 0.001 and momentum 0.9 were used. The early stopping strategy was used with a patience of 10, and cross entropy loss 4 was monitored on the validation set. To perform statistical tests (students ttest), each model was run with 10 different randomstates 5 which were consistent among all experiments, to enable like for like comparisons. The network was trained from scratch without using any pre-trained models, and no augmentation techniques were applied.
3 Results
All the results reported and discussed are on testing sets based on the average of 10 runs with different randomstates.
3.1 Varying Depth with Standard Image Resolution
Figure 1 shows how using shallower versions of the standard ResNet18 architecture affects the classification accuracy on the benchmark datasets used here. Performance generally improves with increasing depth, e.g., OrganSMNIST shows the typical behaviour often reported in the literature. However, other depths appear at least competitive on some datasets, e.g., DermaMNIST, and as shown in Table 4, in one case a slightly shallower CNN outperformed the standard ResNet.
Figure 1:
Table 4:
| Dataset name | Best Depth | Significance |
| OrganAMNIST | Block3 | p < 0.05 * |
3.2 Varying Image Resolution with Standard Depth
Figure 2 shows how using lower image resolutions than the recommended ResNet18 approach affects the classification accuracy on the benchmark datasets used here. Performance often improves or is highest with increasing resolution, e.g., BloodMNIST. However, other resolutions appear at least competitive on some datasets, e.g., PneumoniaMNIST, and as shown in Table 5, in a number of cases a lower image resolution significantly outperforms the recommended standard of 224x224.
Figure 2:
Table 5:
| Dataset name | Resolution | Significance |
| PneumoniaMNIST | 64x64 128x128 | p < 0.01 ** p < 0.05 * |
| OrganAMNIST | 64x64 | p < 0.05 * |
| OrganSMNIST | 64x64 128x128 | p < 0.001 *** p < 0.001 *** |
3.3 Varying Both Image Resolution and Depth
These results show that performance can be significantly improved by decreasing either depth or image resolution from the standard ResNet18. Figure 3 shows how varying both simultaneously can affect the classification accuracy across a wide range of the parameter space on the benchmark datasets. Table 6 details all of the combinations that significantly outperform the recommended standard depth of 4 blocks with 224x224 resolution. The two fundamental aspects of ResNet18, and by implication CNN in general, behave in non-linear ways. The detailed average accuracy across 10 runs for each dataset under each condition is shown in Table 7.
Figure 3:
Table 6:
| Dataset | Resolusion | Block | FMSize | Significance |
| BloodMNIST | 224x224 | 3 | 14x14 | NS |
| BreastMNIST | 64x64 | 3 | 4x4 | NS |
| DermaMNIST | 28x28 | 1 | 7x7 | NS |
| DermaMNIST | 64x64 | 2 | 8x8 | p < 0.05 * |
| DermaMNIST | 128x128 | 3 | 8x8 | NS |
| DermaMNIST | 224x224 | 3 | 14x14 | NS |
| OrganAMNIST | 64x64 | 3 | 4x4 | NS |
| OrganAMNIST | 64x64 | 4 | 2x2 | p < 0.05 * |
| OrganAMNIST | 128x128 | 4 | 4x4 | NS |
| OrganAMNIST | 224x224 | 3 | 14x14 | p < 0.05 * |
| OrganCMNIST | 64x64 | 3 | 4x4 | NS |
| OrganCMNIST | 128x128 | 3 | 8x8 | NS |
| OrganCMNIST | 128x128 | 4 | 4x4 | NS |
| OrganSMNIST | 64x64 | 3 | 4x4 | p < 0.05 * |
| OrganSMNIST | 64x64 | 4 | 2x2 | p < 0.001 *** |
| OrganSMNIST | 128x128 | 4 | 4x4 | p < 0.001*** |
| PneumoniaMNIST | 28x28 | 1 | 7x7 | NS |
| PneumoniaMNIST | 28x28 | 2 | 4x4 | NS |
| PneumoniaMNIST | 28x28 | 3 | 2x2 | NS |
| PneumoniaMNIST | 28x28 | 4 | 1x1 | NS |
| PneumoniaMNIST | 64x64 | 2 | 8x8 | p < 0.01 ** |
| PneumoniaMNIST | 64x64 | 3 | 4x4 | p < 0.01 ** |
| PneumoniaMNIST | 64x64 | 4 | 2x2 | p < 0.01 ** |
| PneumoniaMNIST | 128x128 | 2 | 16x16 | p < 0.05 * |
| PneumoniaMNIST | 128x128 | 3 | 8x8 | p < 0.01 ** |
| PneumoniaMNIST | 128x128 | 4 | 4x4 | p < 0.05 * |
| PneumoniaMNIST | 224x224 | 2 | 28x28 | NS |
| PneumoniaMNIST | 224x224 | 3 | 14x14 | NS |
| RetinaMNIST | 128x128 | 1 | 32x32 | NS |
| RetinaMNIST | 224x224 | 3 | 14x14 | NS |
Table 7:
| Dataset - Resolution | Block1 | Block2 | Block3 | Block4 |
| BloodMNIST - 28x28 | 0.914 | 0.914 | 0.915 | 0.913 |
| BloodMNIST - 64x64 | 0.960 | 0.968 | 0.967 | 0.961 |
| BloodMNIST - 128x128 | 0.961 | 0.975 | 0.978 | 0.973 |
| BloodMNIST - 224x224 | 0.976 | 0.970 | 0.982 | 0.979 |
| BreastMNIST - 28x28 | 0.830 | 0.847 | 0.848 | 0.826 |
| BreastMNIST - 64x64 | 0.814 | 0.824 | 0.855 | 0.851 |
| BreastMNIST - 128x128 | 0.795 | 0.809 | 0.836 | 0.849 |
| BreastMNIST - 224x224 | 0.799 | 0.818 | 0.787 | 0.853 |
| DermaMNIST - 28x28 | 0.751 | 0.731 | 0.732 | 0.738 |
| DermaMNIST - 64x64 | 0.743 | 0.760 | 0.735 | 0.736 |
| DermaMNIST - 128x128 | 0.732 | 0.747 | 0.757 | 0.727 |
| DermaMNIST - 224x224 | 0.725 | 0.734 | 0.753 | 0.748 |
| OrganAMNIST - 28x28 | 0.977 | 0.990 | 0.992 | 0.872 |
| OrganAMNIST - 64x64 | 0.978 | 0.992 | 0.995 | 0.996 |
| OrganAMNIST - 128x128 | 0.961 | 0.987 | 0.994 | 0.995 |
| OrganAMNIST - 224x224 | 0.981 | 0.976 | 0.996 | 0.994 |
| OrganCMNIST - 28x28 | 0.939 | 0.954 | 0.957 | 0.957 |
| OrganCMNIST - 64x64 | 0.935 | 0.962 | 0.972 | 0.967 |
| OrganCMNIST - 128x128 | 0.902 | 0.951 | 0.972 | 0.974 |
| OrganCMNIST - 224x224 | 0.881 | 0.931 | 0.963 | 0.970 |
| OrganSMNIST - 28x28 | 0.842 | 0.878 | 0.892 | 0.892 |
| OrganSMNIST - 64x64 | 0.838 | 0.888 | 0.910 | 0.916 |
| OrganSMNIST - 128x128 | 0.815 | 0.863 | 0.898 | 0.914 |
| OrganSMNIST - 224x224 | 0.789 | 0.842 | 0.881 | 0.901 |
| PneumoniaMNIST - 28x28 | 0.959 | 0.960 | 0.960 | 0.958 |
| PneumoniaMNIST - 64x64 | 0.952 | 0.966 | 0.965 | 0.965 |
| PneumoniaMNIST - 128x128 | 0.955 | 0.966 | 0.965 | 0.965 |
| PneumoniaMNIST - 224x224 | 0.936 | 0.960 | 0.962 | 0.956 |
| RetinaMNIST - 28x28 | 0.515 | 0.502 | 0.477 | 0.484 |
| RetinaMNIST - 64x64 | 0.516 | 0.519 | 0.494 | 0.487 |
| RetinaMNIST - 128x128 | 0.527 | 0.518 | 0.504 | 0.501 |
| RetinaMNIST - 224x224 | 0.521 | 0.520 | 0.525 | 0.522 |
4 Conclusion and Discussion
We have explored two approaches of using neural networks for the classification of biomedical images. We have adopted the commonly used ResNet18 as our baseline and compared it with shallower models using lower-resolution images. The results suggest that the standard depth is robust when using image sizes of 224x224. However, this is not necessarily the case using lower resolutions. Results obtained by simultaneously varying depth with resolution suggest a non-linear behaviour following one of three patterns. The expected one is that performance often increases with depth regardless of resolution (e.g. Figure 3 using OrganCMNIST). However, this is not always the case and sometimes increasing depth and/or resolution provides no significant benefit in classification performance (e.g. Figure 3 using BloodMNIST). In some other examples the performance is sensitive to both the resolution and depth combination (e.g. Figure 3 using PneumoniaMNIST). For the best-performing combinations, the feature maps of the top models vary across different datasets and even within the same dataset (e.g., in Table 6 for OrganAMNIST, the feature map sizes differ significantly, with one being 2x2 and the other 14x14). This suggests that there are multiple ways to extract salient features from the original images through different mathematical approaches.
Our main contribution is a systematic evaluation of network depth and image resolution combinations in medical images, a factor overlooked in related studies, including the recent one by Zhang et al. [13], which did not explore the impact of image resolution. While direct comparison is limited by different data partitions, our best models outperform theirs in accuracy for DermMNIST, OrganSMNIST, PneumoniaMNIST, and several other datasets.
Acknowledgments
The work was funded by a UWE Vice-Chancellor’s Early Career Researcher Development Award. During this study, TB and HL were visiting researchers to N/Lab6. We are grateful to members of both N/Lab and the UWE School of Computing and Creative Technologies for many invaluable discussions. Thanks for Dr. Sue Scarborough for the advice on visualisation.
Footnotes
References
[1]
Giansalvo Cirrincione, Rahul Ranjeev Kumar, Ali Mohammadi, Shahin Hedayati Kia, Pietro Barbiero, and Jacopo Ferretti. 2020. Shallow versus deep neural networks in gear fault diagnosis. IEEE Transactions on Energy Conversion 35, 3 (2020), 1338–1347.
[2]
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition. Ieee, 248–255.
[3]
Tomoya Furukawa and Qianfgu Zhao. 2017. On Extraction of Rules from Deep Learner: The Deeper, the Better?. In 2017 IEEE 15th Intl Conf on Dependable, Autonomic and Secure Computing, 15th Intl Conf on Pervasive Intelligence and Computing, 3rd Intl Conf on Big Data Intelligence and Computing and Cyber Science and Technology Congress (DASC/PiCom/DataCom/CyberSciTech). IEEE, 54–60.
[4]
Ian Goodfellow, Yoshua Bengio, and Aaron Courville. 2016. Deep learning. MIT press.
[5]
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition. 770–778.
[6]
Eran Malach and Shai Shalev-Shwartz. 2019. Is deeper better only when shallow is good? Advances in Neural Information Processing Systems 32 (2019).
[7]
Hrushikesh Mhaskar, Qianli Liao, and Tomaso Poggio. 2016. Learning functions: when is deep better than shallow. arXiv preprint arXiv:https://arXiv.org/abs/1603.00988 (2016).
[8]
ARTZAI Picon Ruiz, AITOR Alvarez Gila, Unai Irusta, JONE Echazarra Huguet, et al. 2020. Why deep learning performs better than classical machine learning? Dyna Ingenieria E Industria (2020).
[9]
Salvador Robles Herrera, Martine Ceberio, and Vladik Kreinovich. 2022. When is deep learning better and when is shallow learning better: Qualitative analysis. International Journal of Parallel, Emergent and Distributed Systems 37, 5 (2022), 589–595.
[10]
Abderrezzaq Sendjasni, David Traparic, and Mohamed-Chaker Larabi. 2022. Investigating Normalization Methods for CNN-Based Image Quality Assessment. In 2022 IEEE International Conference on Image Processing (ICIP). IEEE, 4113–4117.
[11]
Ohad Shamir. 2018. Are resnets provably better than linear predictors? Advances in neural information processing systems 31 (2018).
[12]
Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. 2023. MedMNIST v2-A large-scale lightweight benchmark for 2D and 3D biomedical image classification. Scientific Data 10, 1 (2023), 41.
[13]
Yu Zhang, Pengxing Cai, Yanan Sun, Zhiming Zhang, Zhenyu Lei, and Shangce Gao. 2024. A Lightweight Multi-Dendritic Pyramidal Neuron Model with Neural Plasticity on Image Recognition. IEEE Transactions on Artificial Intelligence (2024).
Index Terms
- ResNet18 Performance: Impact of Network Depth and Image Resolution on Image Classification
Index terms have been assigned to the content through auto-classification.
Recommendations
Attention-aware perceptual enhancement nets for low-resolution image classification
AbstractClassifying low-resolution (LR) images is notoriously challenging because of their noisy representation and limited information. Existing approaches mainly solve this challenge by training carefully designed architectures on LR ...
Blurred Image Recognition: A Joint Motion Deblurring and Classification Loss-Aware Approach
Artificial Neural Networks and Machine Learning – ICANN 2021AbstractImage motion blur can severely affect the performance of the image recognition model. Traditional methods to tackle this problem usually involve image motion deblurring to improve the image quality before its recognition. However, traditional ...
Deep learning for multisensor image resolution enhancement
GeoAI '17: Proceedings of the 1st Workshop on Artificial Intelligence and Deep Learning for Geographic Knowledge DiscoveryWe describe a deep learning convolutional neural network (CNN) for enhancing low resolution multispectral satellite imagery without the use of a panchromatic image. For training, low resolution images are used as input and corresponding high resolution ...
Comments
Please enable JavaScript to view thecomments powered by Disqus.Information & Contributors
Information
Published In
October 2024
373 pages
ISBN:9798400718014
DOI:10.1145/3704137
Copyright © 2024 Copyright held by the owner/author(s).
This work is licensed under a Creative Commons Attribution 4.0 International License.
Publisher
Association for Computing Machinery
New York, NY, United States
Publication History
Published: 03 March 2025
Check for updates
Author Tags
Qualifiers
- Research-article
Conference
ICAAI 2024
ICAAI 2024: 2024 The 8th International Conference on Advances in Artificial Intelligence
October 17 - 19, 2024
London, United Kingdom
Contributors
Other Metrics
Bibliometrics & Citations
Bibliometrics
Article Metrics
- 0Total Citations
- 97Total Downloads
- Downloads (Last 12 months)97
- Downloads (Last 6 weeks)97
Reflects downloads up to 08 Mar 2025
Other Metrics
Citations
View Options
Login options
Check if you have access through your login credentials or your institution to get full access on this article.
Sign in