Ensemble Cross UNet Transformers for Augmentation of Atomic Electron Tomography | IEEE Journals & Magazine | IEEE Xplore

Ensemble Cross UNet Transformers for Augmentation of Atomic Electron Tomography


Abstract:

Investigating atomic structures and tracing precise atomic coordinates through atomic electron tomography (AET) are crucial for understanding the properties of functional...Show More

Abstract:

Investigating atomic structures and tracing precise atomic coordinates through atomic electron tomography (AET) are crucial for understanding the properties of functional materials. Unfortunately, the fidelity of reconstructed 3-D tomograms is frequently compromised due to geometric constraints, low radiation doses, and tile angle errors, leading to the presence of unwanted artifacts. A previously established solution involves using a 3D-UNet methodology to address imperfections within tomograms, effectively reducing a significant portion of these artifacts. However, residual artifacts persist in the augmented tomograms, along with the omission of surface details, which subsequently hinders atomic measurement tasks, including precise atom tracing. This article presents an innovative augmentation architecture termed ensemble cross UNet Transformers (EC-UNETRs), designed to systematically remove artifacts and meticulously recover missing details. Our ensemble cross (EC) network can be regarded as an assemblage of hierarchical subnetworks that furnish refined fine-grained features, which yield distinct outputs along both the horizontal and vertical axes. Moreover, we present a skip fusing unit designed to enhance the stability and flexibility of the subnetworks, thereby facilitating the seamless integration of features across diverse architectural levels. In order to efficiently capture long-range spatial dependencies, we incorporate Transformer-based efficient paired attentions into both the encoder and decoder stages of the architecture. The application of EC-UNETR has yielded remarkable outcomes in the augmentation of both simulated and real 3-D tomograms, leading to a notable enhancement in tomographic quality. Specifically, a 45.2% reduction in root-mean-square error (RMSE) has been achieved compared with baseline, coupled with a noteworthy 5.5% decrease when contrasted with the state-of-the-art 3-D tomogram analysis methodology. The code is available at https://g...
Article Sequence Number: 5021714
Date of Publication: 20 May 2024

ISSN Information:

Funding Agency:


References

References is not available for this document.