Preliminary Result: AI-Generated Neutrophil Image using Deep Convolution GAN for Data Augmentation.
GANs (Generative Adversarial Networks) have grown impressively, producing significant photorealistic visuals that imitate the content of datasets they were trained to replicate. A GAN is essentially two neural networks that feed into one other. One generates increasingly accurate data, while the oth...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Conference or Workshop Item |
| Language: | English English |
| Published: |
2022
|
| Subjects: | |
| Online Access: | http://irep.iium.edu.my/98826/1/International%20Islamic%20University%20Malaysia%20Mail%20-%20Fwd_%20Confirmation%20of%20Acceptance%20%28Oral%29.pdf http://irep.iium.edu.my/98826/2/ICBME%20Presentation.pdf http://irep.iium.edu.my/98826/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | GANs (Generative Adversarial Networks) have grown impressively, producing significant photorealistic visuals that imitate the content of datasets they were trained to replicate. A GAN is essentially two neural networks that feed into one other. One generates increasingly accurate data, while the other increases its capacity to classify such data over time. One recurring subject in medical imaging is whether GANs can be as effective at producing usable medical data as they are at producing realistic images. The deep learning model is data-hungry in nature, it requires a lot of example images to train well. However, due to the lack of medical images, data augmentation comes in handy to generate extra medical images using GAN. This paper aims to generate microscopic
peripheral blood cell images, specifically neutrophils, as a form of data augmentation to optimize haematological diagnosis. To accomplish this, we developed a Deep Convolution GAN (DCGAN) and trained with 3329 neutrophil images. For the preliminary result, we will present our work on the impact of different learning rates and optimizers of DCGAN on the generated images and training losses. The quality of the generated images is far from perfect from the dataset we want to imitate, also the convergence of the model is slow and not stable. Yet, there were reasonably generated images during the training where the model has a rough idea about the neutrophil structure. |
|---|