Click here to

Session: Multi-Disciplinary: Imaging for Treatment Planning [Return to Session]

Self-Supervised Learning Based CT Imaging with High Through-Plane Resolution

H Xie1, Y Lei2, T Wang1,3, Z Tian1,3, J Roper1,3, J Bradley1,3, X Tang2,3, T Liu1,3, X Yang1,3*, (1) Department of Radiation Oncology, (2) Department of Radiology and Imaging Sciences, and (3) Winship Cancer Institute, Emory University, Atlanta, GA


SU-IePD-TRACK 3-5 (Sunday, 7/25/2021) 5:30 PM - 6:00 PM [Eastern Time (GMT-4)]

Purpose: CT images for radiation therapy planning are usually acquired in thick slice to reduce imaging dose, especially for pediatric patients. However, this practice may degrade the quality of dose calculation and target delineation. In this work, a self-supervised deep learning workflow is proposed to synthesize high through-plane (sagittal and coronal planes) resolution CT images without requiring ground truth images for training, as patient images of thin slices are often not available for many tumor sites.

Methods: The proposed workflow was designed to facilitate the convolutional neural networks to learn the mapping from low resolution (LR) to high resolution (HR) images in the in-plane (axial) direction. In the workflow, the through-plane LR CT images were sliced along the sagittal and coronal planes and fed through two parallel data flow paths. During the inference step, the through-plane HR images were generated by feeding the respective LR sagittal and coronal images to the trained networks.

Results: The qualitative results show the capability of the proposed method for generating high through-plane resolution with data of 75 head and neck and 20 lung cancer patients’ CT images. The image quality metrics can be all enhanced with the proposed method by from ~7% for structural similarity index measurement (SSIM) to ~130% for edge keeping index (EKI). All the improvements of the measure metrics are confirmed to be statistically significant with paired two-sample t-test analysis (p=0.014 for SSIM and p<0.001 for EKI).

Conclusion: The proposed deep learning method for CT image generation with high through-plane resolution is self-supervised, which means it does not rely on ground truth for network training. In addition, the assumption that the in-plane HR information can supervise the through-plane HR generation is confirmed and anticipated to potentially inspire more researches on this topic to further improve the through-plane resolution of medical images.

Funding Support, Disclosures, and Conflict of Interest: This research is supported in part by the National Cancer Institute of the National Institutes of Health under Award Number R01CA215718, and Pilot Grant by the Winship Cancer Institute of Emory University.



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email