🤖 AI Summary
Cryo-electron tomography (cryo-ET) suffers from a critical scarcity of realistic, annotated training data: deep learning methods require large-scale ground-truth annotations, while physics-based simulation is computationally expensive and struggles to model complex, biologically realistic degradations. To address this, we propose the first neural style transfer framework tailored for cryo-ET simulation—a fully end-to-end 3D style transfer method that transforms synthetic or low-fidelity volume data into high-fidelity, biophysically consistent tomograms without paired supervision. Our approach integrates Adaptive Instance Normalization (AdaIN) with a 3D U-Net architecture, augmented by frequency-domain constraints, structural similarity (SSIM) loss, and a biologically inspired contrast prior. It further enables controllable artifact injection and multi-scale structural preservation. On multiple benchmarks, our simulated tomograms achieve an 18.7% SSIM improvement and yield a 12.3% Dice score gain in downstream segmentation, significantly enhancing pretraining efficacy and robustness evaluation of cryo-ET analysis models.