Dual residual learning of frequency fingerprints in detecting synthesized biomedical imagery

Abstract

Artificial synthesis of biomedical imagery is an evolving threat yet under-addressed. The integrity of medical imaging is important for accurate diagnosis and treatment. This study addresses the potential threat of fabricated biomedical imagery, focusing on synthetic dermatological lesions and CT nodules. The Representation Similarity Matrix measured the quantitative authenticity to account for similarities of synthesized data with authentic data. The study explores traces of manipulation from frequency signatures of synthesized imagery. We propose a novel combinatorial architecture, the Dual Residual Network (DRN), capturing hidden residual traces from low-frequency fingerprints of synthetic data and exposing hidden forgeries. DRN achieves near-perfect detection rates with an accuracy of 98.80% for CT nodules and 98.97% for lesions. Equal Error Rates of the model on the two datasets exhibited a marginal improvement of 57.87% in the CT nodules compared to the skin lesions. Sensitivity and specificity play a significant role in medical diagnostics. The model achieved sensitivities of 99.31% and 98.45% and specificity of 98.80% and 99.60% for each dataset, respectively. Further verification of the frequency traces was performed by analyzing gradients in the target concepts that led to decision-making. This study equips the medical field with a powerful tool to combat the evolving threat of synthetic fraud, safeguarding patient and client safety. The potential of the technique extends beyond healthcare, offering a blueprint for tackling synthetic data across diverse domains.