AI-driven Single Image Super Resolution for Improved Neuron Segmentation

Abstract

Connectomics research relies heavily on high-resolution imaging of neurons, allowing for the segmentation and tracing of nerve structures. Typically, high-resolution images cover only limited areas, while broader overviews are captured at lower resolutions. As segmentation techniques continue to improve, the analysis of these low-resolution regions has become of increasing interest. AI-driven super resolution models offer the potential to upsample low-resolution neural images, enabling automated segmentation in regions that were previously unusable for analysis and increasing precision in areas with high feature density. As this specific application of super resolution is previously unexplored, and given the growing variety of model architectures available, this work investigates three representative models of different architectures, Real-ESRGAN, SR3, and EDT, and variety of training loss functions. The goal is to compare the strengths and limitations of these architectures when fine-tuned on serial block-face electron microscopy images. This task demands a high degree of structural consistency between low and high resolution outputs. REAL-ESRGAN and SR3 were found to be prone to hallucinations and artifacts, which can hinder downstream applications. In contrast, the fine-tuned EDT models tended to produce overly smooth outputs and in this removed small features. Some improvement was achieved with a task-specific EDT model trained from the ground up and the use of structural similarity–based loss functions.