Synthesisation of diverse and photorealistic lung images for lung diseases recognition using deep learning techniques
Deep Learning (DL) techniques leverage the rich data diversity in the modern era of Big Data, allowing possibilities of achieving human-level performances in many real-world problems. However, newly emerged problems such as the ongoing COVID-19 pandemic have underscored the challenge of acquiring a...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English |
| Published: |
2023
|
| Subjects: | |
| Online Access: | https://eprints.ums.edu.my/id/eprint/40749/1/24%20PAGES..pdf https://eprints.ums.edu.my/id/eprint/40749/2/FULLTEXT..pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Deep Learning (DL) techniques leverage the rich data diversity in the modern era of Big Data, allowing possibilities of achieving human-level performances in many real-world problems. However, newly emerged problems such as the ongoing COVID-19 pandemic have underscored the challenge of acquiring a larger corpus of data during the early stages of the pandemic due to the availability of samples, data protection policies, labour, and facility resources. Furthermore, the effectiveness of existing DL models that are trained on older datasets might be vulnerable to the continuous emergence of COVID-19 variants that may potentially result in distribution shifts. This research introduces a data synthesisation framework named stacked residual dropout generative adversarial network (sRD-GAN), which alleviates the problem of data paucity by generating synthetic lung medical images that contain precise radiographic feature annotations. The underlying design of sRD-GAN is an Image-to-Image translation setting that facilitates instance-level diversity via the latent space stochasticity induced by the novel stacked residual dropout (sRD) regularization. To this end, experiments show that sRD-GAN achieved perceptually significant structural dissimilarities of the ground glass opacities (GGO) from diverse COVID-19 CT images without disentangling the content-style attributes of the images as in conventional multimodal image translation techniques. Since the sRD regularization is a strategic incorporation of the conventional dropout regularization, which can be generalized across neural network models, the sRD regularization can be easily incorporated into existing image synthesizer models without modifying the original setup of these models. In addition, a new training loss function known as adaptive pixel consistency loss is proposed for effective noise reduction by encouraging structural similarity of the invariance features of the images from both domains. Quantitative results show that the synthetic COVID-19 CT images achieve a promising Fréchet Inception Distance (FID) of 58.68, which is superior to existing GAN baselines such as GAN (157.18), CycleGAN (115.14), and One-to-one CycleGAN (94.11). Visual examination of the synthetic images also indicates excellent perceptual image quality and realism, where synthetic radiographic features of GGO achieve consistency with real COVID-19 CT images examined by an experienced radiologist. Furthermore, the effectiveness of the proposed sRD-GAN is also validated on Community-Acquired Pneumonia (CAP) CT images and COVID-19 X-Ray images, which achieved comparable performances with COVID-19 CT images. This suggests that the proposed method can be easily extended to other similar applications. Lastly, the sRD-GAN is applied to the problem of COVID-19 disease recognition in the form of dynamic data augmentation. Empirical results suggest that synthetic images can approximate real data distribution for model training purposes. Specifically, the VGG19 models achieve the highest accuracy score at 97.54% on the test set when training with fully synthetic COVID-19 CT images in 3000-images dataset size, contributing to a 12.95% accuracy improvement from training with only real image data. |
|---|