Image Imputation with conditional generative adversarial networks captures clinically relevant imaging features on...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Kidney cancer is a common adult malignancy often diagnosed using four-phase computed tomography (CT) imaging.
• Missing or corrupted CT images pose a significant challenge in kidney cancer detection, diagnosis, and treatment planning.
• Deep learning, specifically conditional generative adversarial networks (cGANs), shows promise for imputing missing medical imaging data.

Purpose of the Study:

• To explore the clinical utility of images imputed by cGANs for four-phase kidney cancer CT studies.
• To assess the accuracy and reliability of cGAN-imputed images in retaining clinically relevant diagnostic information.

Main Methods:

• A cGAN was trained on 333 patient datasets to impute a missing CT phase given the other three.
• The cGAN's imputation performance was evaluated on a separate test set of 37 patients.
• Clinically relevant imaging features (21 total, including 13 categorical) were manually extracted from both true and imputed images for comparison.

Main Results:

• Imputed images demonstrated high agreement with true images for all 13 categorical clinical features (over 85% agreement).
• This accuracy was consistent across different imaging phases.
• Imputed images showed good agreement in quantitative radiomic features like mean intensity and enhancement, and maintained diagnostic characteristics for benign or malignant classification.

Conclusions:

• cGAN-imputed images effectively retain essential qualitative and quantitative features from four-phase kidney cancer CT scans.
• The high accuracy and feature retention suggest significant potential for clinical application of cGAN-imputed images.
• This technology could help overcome challenges posed by missing or corrupted data in medical imaging, improving patient care.