Heterogenous Dimensional Super Resolution of 3D CT Scans Using Transformers

Authors: Helen Zhang

Abstract:

Accurate segmentation of the airways from CT scans is crucial for early diagnosis of lung cancer. However, the existing airway segmentation algorithms often rely on thin-slice CT scans, which can be inconvenient and costly. This paper presents a set of machine learning-based 3D super-resolution algorithms along heterogenous dimensions to improve the resolution of thicker CT scans to reduce the reliance on thin-slice scans. To evaluate the efficacy of the super-resolution algorithms, quantitative assessments using PSNR (Peak Signal to Noise Ratio) and SSIM (Structural SIMilarity index) were performed. The impact of super-resolution on airway segmentation accuracy is also studied. The proposed approach has the potential to make airway segmentation more accessible and affordable, thereby facilitating early diagnosis and treatment of lung cancer.

Keywords: 3D super-resolution, airway segmentation, thin-slice CT scans, machine learning.

References:

[1] A. Kudo, et. al. Virtual Thin Slice: 3D Conditional GAN-based Super-resolution for CT Slice Interval. https://arxiv.org/pdf/1908.11506.pdf
[2] S. Park, et. al. Deep Learning Algorithm for Reducing CT Slice Thickness: Effect on Reproducibility of Radiomic Features in Lung Cancer. Korean J Radiol. 2019 Oct; 20(10): 1431–1440.
[3] U. Agrawal, et. al. Enhancing Z-resolution in CT volumes with deep residual learning. SPIE Medical Imaging 2021.
[4] H. Xie, et. al. High through-plane resolution CT imaging with self-supervised deep learning. 2021 Phys. Med. Biol. 66 145013.
[5] W. Bae, et. al. Residual CNN-based Image Super-Resolution for CT Slice Thickness Reduction using Paired CT Scans: Preliminary Validation Study. https://openreview.net/forum?id=S1RzBW2oz.
[6] M. Kiss, et. al. Z-Super Resolution CT-Image Generation with A Deep 3D Fully Convolutional Neural Network. https://doi.org/10.1016/j.ijrobp.2020.07.249.
[7] C. You, et. al. CT Super-resolution GAN Constrained by the Identical, Residual, and Cycle Learning Ensemble (GAN-CIRCLE) https://arxiv.org/pdf/1808.04256.pdf.
[8] J. Liang, et. al. SwinIR: Image Restoration Using Swin Transformer. https://arxiv.org/abs/2108.10257.
[9] https://github.com/Project-MONAI
[10] NLST dataset in the Cancer Image Archive. https://cdas.cancer.gov/datasets/nlst/