Self-Supervised Adversarial Learning with a Limited Dataset for Electronic Cleansing in Computed Tomographic...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Radiology

Background:

• Current electronic cleansing (EC) methods for computed tomographic colonography (CTC) rely on image segmentation, limiting accuracy to voxel resolution.
• Traditional deep learning approaches for EC are constrained by the limited availability of annotated CTC datasets for training.

Purpose of the Study:

• To evaluate the technical feasibility of a novel self-supervised adversarial learning scheme for EC in CTC.
• To achieve subvoxel accuracy in EC using a limited training dataset.

Main Methods:

• A three-dimensional generative adversarial network (3D GAN) was pre-trained on CTC phantom data for EC.
• The 3D GAN was fine-tuned using a self-supervised scheme on a case-by-case basis.
• The 3D GAN architecture was optimized through phantom studies.

Main Results:

• The self-supervised 3D GAN achieved subvoxel accuracy for EC.
• Visual quality of virtual cleansing by the 3D GAN favorably compared to commercial EC software.
• The method demonstrated effectiveness on 18 clinical CTC cases.

Conclusions:

• The proposed self-supervised 3D GAN is a potentially effective approach for EC in CTC.
• This method addresses technical challenges by enabling training on small, unannotated datasets with subvoxel accuracy.