Enhanced Electron Beam Tomography Reconstruction via Adaptive Multi-Modal Neural Network Fusion

**Abstract:** This paper details a novel framework for enhancing electron beam tomography (EBT) image reconstruction, specifically addressing the limitations of current iterative reconstruction algorithms in handling low-dose data and artifacts arising from beam-sample interactions. Our approach, termed Adaptive Multi-Modal Neural Network Fusion (AMMNF), seamlessly integrates data from multiple modalities (transmission data, scattered electron data, and probe current data) using a deep neural network architecture. This fusion process is guided by adaptive weighting based on real-time image quality assessment of intermediate reconstruction steps, significantly improving image resolution, contrast, and reducing noise compared to conventional methods, while maintaining clinically relevant acquisition times. We demonstrate the potential for translating this improved resolution and diagnostic accuracy to early-stage disease detection in biological samples, estimated to impact the medical imaging market by 5-7% within 5 years through faster diagnosis and reduced patient exposure.