Synthetic procedural noise and neural networks: enhancing biomedical images with purely artificial data

Neural networks (NNs) are capable of generalizing beyond training data, yet the required level of similarity between the training data and inference data remains unclear. This research investigates the potential of training NNs with purely synthetic data to achieve effective generalization in biomedical applications. Specifically, Perlin noise, a procedurally generated noise, was used to simulate pseudo-random signals with local spatial correlations resembling that of biomedical data. The study trained a U-net autoencoder on purely synthetic data to denoise time-domain full-field optical coherence tomography (TD-FF-OCT) images of human corneas. Remarkably, the NN successfully denoised real TD-FF-OCT images, suppressing not only the Gaussian noise but also the correlated interference fringes. We explore this method for denoising, dehazing and contrast enhancement across various imaging modalities and in particular in cases when clean data is fundamentally unavailable. Finally, we compare this method to other state-of-the-art synthetic data tools such as diffusion models.