The Effect of Image Resolution on Deep Learning in Radiography

Abstract

Purpose: To examine variations of convolutional neural network (CNN) performance for multiple chest radiograph diagnoses and image resolutions.

Materials and methods: This retrospective study examined CNN performance using the publicly available National Institutes of Health chest radiograph dataset comprising 112 120 chest radiographic images from 30 805 patients. The network architectures examined included ResNet34 and DenseNet121. Image resolutions ranging from 32 × 32 to 600 × 600 pixels were investigated. Network training paradigms used 80% of samples for training and 20% for validation. CNN performance was evaluated based on area under the receiver operating characteristic curve (AUC) and label accuracy. Binary output networks were trained separately for each label or diagnosis under consideration.

Results: Maximum AUCs were achieved at image resolutions between 256 × 256 and 448 × 448 pixels for binary decision networks targeting emphysema, cardiomegaly, hernias, edema, effusions, atelectasis, masses, and nodules. When comparing performance between networks that utilize lower resolution (64 × 64 pixels) versus higher (320 × 320 pixels) resolution inputs, emphysema, cardiomegaly, hernia, and pulmonary nodule detection had the highest fractional improvements in AUC at higher image resolutions. Specifically, pulmonary nodule detection had an AUC performance ratio of 80.7% ± 1.5 (standard deviation) (0.689 of 0.854) whereas thoracic mass detection had an AUC ratio of 86.7% ± 1.2 (0.767 of 0.886) for these image resolutions.

Conclusion: Increasing image resolution for CNN training often has a trade-off with the maximum possible batch size, yet optimal selection of image resolution has the potential for further increasing neural network performance for various radiology-based machine learning tasks. Furthermore, identifying diagnosis-specific tasks that require relatively higher image resolution can potentially provide insight into the relative difficulty of identifying different radiology findings. Supplemental material is available for this article. © RSNA, 2020See also the commentary by Lakhani in this issue.

Figure 1:

Comparison of chest radiographs at different image resolutions for patient 103 (60-year-old man with a thoracic mass). The mass finding is visible in all images but with visually observable improved clarity in the higher resolution examples (bottom row).

Figure 2:

Validation set area under the receiver operating characteristic curve (AUROC) for six different diagnostic labels shows improved performance with increased image resolution and a plateau effect on performance improvement for resolutions higher than 224 × 224 pixels. Models were trained with ResNet34 architecture for three subsample epochs. Resolutions shown are as follows: 32 × 32, 64 × 64, 128 × 128, 224 × 224, 256 × 256, 320 × 320, 448 × 448, 512 × 512, and 600 × 600 pixels. Error bars represent standard deviation of the area under the curve calculated via the DeLong method.

Figure 3:

Bar graph shows percentage area under the receiver operating characteristic curve (AUROC) achievable with low-resolution models compared with a higher resolution 320 × 320-pixel resolution model for eight example diagnostic labels. Edema prediction models at 32 × 32-pixel resolution are able to capture the highest percentage of a 320 × 320-pixel resolution model.

Figure 4:

Comparison of area under the receiver operating characteristic curve (AUROC) as a function of input image resolution for “mass” and “nodule” detection models trained for two different architectures (ResNet34 and DenseNet121) for 10 subsample epochs. Mass prediction models achieve better performance at lower resolutions, which is presumptively attributable to the larger size (>3 cm) of pulmonary masses compared with pulmonary nodules.