Sensitivity analysis of multi-defocus features using spatial attention mechanism for optimizing optical measurement in TSOM

Through-focus scanning optical microscopy (TSOM) is a nanometer measurement technology that records the light intensity transform mode, and has been strongly encouraged by the IEEE International Roadmap for Devices and Systems (IRDS): Metrology. TSOM simultaneously collects focused and defocused scattering information to avoid the optical diffraction limit. While recent advancements have improved TSOM precision, challenges persist in extracting valuable features from multi-defocus image sequences, which impedes comprehensive sensitivity analysis essential for optimizing measurement robustness. A sensitivity analysis method for defocused image features based on a spatial attention mechanism is proposed in this paper to evaluate the impact of different defocused positions and other factors on measurement accuracy. The proposed method strategically employs trainable attention layers to dynamically assign region-specific weights across multi-defocus image stacks, enabling the model to prioritize spatially informative patterns while suppressing noise-contaminated areas. Furthermore, by correlating attention maps with sensitivity metrics, the proposed method unifies optical physics with deep learning interpretability. The proposed method adopts spatial attention mechanisms as its architectural core, synergistically integrated with auxiliary components to strengthen critical information extraction while enabling systematic sensitivity analysis through dynamic region-weighting. We have provided a detailed introduction to the design principles and performance of this method on TSOM, and it is expected to be applied to other optical measurement methods.