Pengembangan Metode Interpolasi Temporal Berbasis Deep Learning untuk Akuisisi Data Neutron Computed Tomography

Abstract

Neutron Computed Tomography is a non-destructive imaging technique with high sensitivity to light elements embedded within dense materials. However, its implementation is constrained by long acquisition times caused by low neutron flux, which requires a large number of projection angles to achieve accurate tomographic reconstruction. This limitation decreases experimental efficiency and increases beam time consumption. Therefore, an efficient computational approach is needed to reduce the number of acquired projections while maintaining image quality. This study proposes a temporal interpolation method in the neutron radiographic projection domain using a hybrid Generative Adversarial Network–Long Short-Term Memory (GAN–LSTM) model. The method aims to predict intermediate projection images between acquired angles by learning both spatial and temporal dependencies among projection sequences. The dataset consists of 361 neutron radiography projection images of a 100 cc motorcycle engine block, acquired at the RSG-GAS Neutron Radiography Facility over a 0°–360° angular range with 1° increments. The proposed architecture utilizes an STUNet–TimeConditioned generator to extract spatial features, while LSTM and time-conditioning mechanisms are employed to capture temporal relationships between adjacent projection angles. In addition, a PatchGAN discriminator is used to preserve texture consistency and improve image realism. The model is trained using an input–target interpolation scheme with an angular interval of Δθ = 2°. Experimental results demonstrate that the proposed model successfully generates interpolated projections with high similarity to the ground-truth images. The best-performing model achieves a Structural Similarity Index Measure (SSIM) of 0.9546 and a Peak Signal-to-Noise Ratio (PSNR) of 37.36 dB, indicating strong preservation of spatial structures and minimal reconstruction artifacts. These findings suggest that GAN–LSTM-based temporal interpolation is effective for reducing the number of projection angles required in neutron imaging experiments, thereby improving the efficiency of neutron computed tomography acquisition processes.