Improvement of Fine Structure Preservation in Denoising by Self-Supervised Learning Compared to Supervised Learning

Inline computed tomography (CT) enables rapid quality control but suffers from high noise due to short scan times. While deep learning offers powerful denoising capabilities, standard supervised models require clean reference data which is often unavailable in industrial settings and which tend to overfit training statistics, risking the suppression of anomalous features like defects. This paper investigates the generalization capabilities of self-supervised learning for preserving such unseen fine structures. We conduct a rigorous comparison between supervised (Noise2Clean) and self-supervised (Noise2Noise) denoising, utilizing the ReTWIN strategy to generate independent training pairs from single projection datasets. We derive a theoretical framework linking the gradient variance in Noise2Noise to implicit regularization that favors "flatter" minima, thereby enhancing generalization. Building on this, we propose Noise2Noise Edge-Enhanced (N2NEE), a novel method that modifies the training target with a spectral boost factor to amplify the gradient signal for spatially coherent high-frequency structures. Performance is evaluated via Modulation Transfer Function (MTF) analysis, simulation of artificial pores (1–7 voxels), and application to real-world aluminium casting scans. Quantitative analysis demonstrates that N2NEE achieves a spatial resolution (MTF50) of 14.518 lp/mm, nearly tripling the performance of supervised baselines (5.747 lp/mm). In controlled experiments, N2NEE successfully preserved 1-voxel pores that were obliterated by supervised models, while also retaining the geometric integrity of larger defects. Validation on industrial data confirmed that N2NEE significantly improves the contrast-to-noise ratio of natural porosity without generating artifacts. These results establish N2NEE as a robust, practical solution for inline CT metrology, eliminating the need for ground truth data while ensuring reliable defect detection.