Interpretable Similarity of Synthetic Image Utility

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Synthetic medical image data is crucial for developing privacy-preserving deep learning (DL) clinical decision support (CDS) systems.
• Current methods for assessing synthetic image similarity lack quantitative interpretability.
• A key research gap exists in evaluating the utility of synthetic data for specific medical applications.

Purpose of the Study:

• To propose a novel, interpretable measure for quantitatively assessing the similarity between synthetic and real medical image datasets.
• To evaluate the utility of synthetic data for developing DL-based CDS systems.
• To provide a method that explains why certain synthetic datasets are more useful than others for clinical applications.

Main Methods:

• Development of Interpretable Utility Similarity (IUS), inspired by generalized neural additive models.
• Comparison of IUS with existing measures like inception-based metrics and classification performance.
• Experimental validation on diverse medical imaging modalities (endoscopic, dermoscopic, fundus, X-ray, ultrasound).

Main Results:

• IUS provides an interpretable assessment of synthetic image utility for DL-based CDS.
• Selecting synthetic images with high IUS demonstrated relative classification performance improvements up to 54.6%.
• IUS proved effective across various color and grayscale medical imaging modalities.

Conclusions:

• IUS offers a valuable, interpretable tool for quantitative assessment of synthetic medical image data.
• The proposed measure enhances the development of effective DL-based CDS systems by optimizing synthetic data selection.
• IUS facilitates the creation of more reliable and performant AI tools in healthcare through improved synthetic data utilization.