Can Synthetic Images Improve CNN Performance in Wound Image Classification?

Area of Science:

• Medical Artificial Intelligence
• Computer Vision in Healthcare
• Machine Learning for Medical Imaging

Background:

• Clinical relevance of AI systems depends on high performance, requiring extensive labeled training data.
• Generative Adversarial Networks (GANs) are utilized to synthesize artificial training images for data augmentation when data is scarce.

Purpose of the Study:

• To evaluate the quality of synthetic wound images generated by GANs.
• To assess the impact of synthetic wound images on Convolutional Neural Network (CNN) classification performance.
• To determine the perceived realism of synthetic wound images by clinical experts.

Main Methods:

• Investigated synthetic wound image quality through two primary aspects: CNN classification improvement and expert realism assessment.
• Utilized a dataset of synthetic wound images generated via GANs to augment training data for a CNN.
• Conducted a study with 217 clinical experts to evaluate the realism of the synthetic images.

Main Results:

• A slight improvement in wound-type classification was observed when using GAN-synthesized data.
• The relationship between the size of the synthetic dataset and classification performance remains unclear.
• Clinical experts identified synthetic images as real in only 31% of cases, despite the high visual realism of GAN-generated images.

Conclusions:

• Synthetic wound images generated by GANs offer a slight benefit for CNN-based wound classification.
• Image quality, rather than solely data size, appears to be a more critical factor in enhancing CNN classification accuracy.
• Further research is needed to optimize GANs for generating highly realistic medical images that are indistinguishable from real ones to experts.