The Effect of Image Resolution on Deep Learning in Radiography

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Convolutional Neural Networks (CNNs) are increasingly used for medical image analysis.
• Chest radiography is a common diagnostic tool, and AI can aid in interpreting these images.
• Image resolution is a critical parameter that can influence the performance of deep learning models.

Purpose of the Study:

• To investigate how different image resolutions affect CNN performance for various chest radiograph diagnoses.
• To compare the performance of CNNs (ResNet34, DenseNet121) across a range of image resolutions.
• To identify optimal image resolutions for specific radiological findings.

Main Methods:

• Retrospective analysis of 112,120 chest radiographs from 30,805 patients.
• Utilized ResNet34 and DenseNet121 architectures with image resolutions from 32x32 to 600x600 pixels.
• Evaluated performance using Area Under the Curve (AUC) and label accuracy for binary classification tasks.

Main Results:

• Optimal AUCs were achieved with resolutions between 256x256 and 448x448 pixels.
• Higher image resolutions (320x320 vs. 64x64) significantly improved AUC for emphysema, cardiomegaly, hernia, and nodule detection.
• Pulmonary nodule and thoracic mass detection showed substantial AUC improvements with higher resolution inputs.

Conclusions:

• Selecting appropriate image resolution is crucial for enhancing CNN performance in radiology.
• Higher resolutions can improve the detection of specific findings, though may impact batch size.
• Diagnosis-specific resolution requirements offer insights into the complexity of identifying radiological features.