Diverse COVID‑19 CT image‑to‑image translation with stacked residual dropout
Machine learning models are renowned for their high dependency on a large corpus of data in solving real‑world problems, including the recent COVID‑19 pandemic. In practice, data acquisition is an onerous process, especially in medical applications, due to lack of data availabil‑ ity for newly emerg...
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English English |
| Published: |
MDPI
2022
|
| Subjects: | |
| Online Access: | https://eprints.ums.edu.my/id/eprint/42278/1/ABSTRACT.pdf https://eprints.ums.edu.my/id/eprint/42278/2/FULL%20TEXT.pdf |
| Summary: | Machine learning models are renowned for their high dependency on a large corpus of data in solving real‑world problems, including the recent COVID‑19 pandemic. In practice, data acquisition is an onerous process, especially in medical applications, due to lack of data availabil‑ ity for newly emerged diseases and privacy concerns. This study introduces a data synthesization framework (sRD‑GAN) that generates synthetic COVID‑19 CT images using a novel stacked‑residual dropout mechanism (sRD). sRD‑GAN aims to alleviate the problem of data paucity by generating synthetic lung medical images that contain precise radiographic annotations. The sRD mechanism is designed using a regularization‑based strategy to facilitate perceptually significant instance‑level di‑ versity without content‑style attribute disentanglement. Extensive experiments show that sRD‑GAN can generate exceptional perceptual realism on COVID‑19 CT images examined by an experiment radiologist, with an outstanding Fréchet Inception Distance (FID) of 58.68 and Learned Perceptual Image Patch Similarity (LPIPS) of 0.1370 on the test set. In a benchmarking experiment, sRD‑GAN shows superior performance compared to GAN, CycleGAN, and one‑to‑one CycleGAN. The encour‑ aging results achieved by sRD‑GAN in different clinical cases, such as community‑acquired pneu‑ monia CT images and COVID‑19 in X‑ray images, suggest that the proposed method can be easily extended to other similar image synthetization problems. |
|---|