Effects of data count and image scaling on Deep Learning training.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel at image feature extraction.
• Image interpolation methods can potentially enhance CNN feature extraction.
• The effectiveness of interpolation for data augmentation in CNNs, particularly with limited medical image data, requires investigation.

Purpose of the Study:

• To evaluate the impact of image size augmentation using interpolation on CNN performance.
• To compare interpolation-based augmentation with traditional methods like rotation and inversion.
• To determine the utility of interpolation for CNN training with medical images.

Main Methods:

• Chest radiographs from the Gender01 dataset were used for sex classification.
• Images were enlarged two- and four-fold using the Bilinear interpolation method.
• Performance was compared against data augmentation via image inversion and rotation.

Main Results:

• Image size expansion via interpolation consistently improved average classification accuracy.
• A four-fold increase in image size yielded the most significant accuracy improvement (0.563 to 0.715) with 100 training samples.
• The Bilinear method outperformed rotation/inversion, showing accuracy gains of 0.095 (100 data) and 0.015 (50,000 data).

Conclusions:

• Augmenting image size through interpolation is a beneficial technique for CNN training.
• This method demonstrates stable and high classification accuracy for medical images like chest X-rays.
• Further research is needed to explore the importance of image size across diverse medical imaging datasets.