URECA - Deep Transfer Learning For Abnormality Detection

URECA stands for Undergraduate Research Experience on CAmpus (URECA). It is a 1 year research programme at Nanyang Technological University. I am honoured that my paper was selected to be published in the 4th International Conference on Crowd Science and Engineering (ICCSE 2019). The link to the published paper can be found here.

Data Sources:

• MNIST Dataset
• Stanford Musculoskeletal Radiographs (MURA) Dataset

Implementation of Research

• Conditional Deep Convolutional Generative Adversarial Network (DCGAN) Implementation (MNIST)
• Evaluation of Conditional DCGAN on MNIST
• Conditional DCGAN implementation (MURA)

Results of Research (MNIST)

Epochs	Transfer Learning Model	Data Size	Accuracy
10	MobileNet	320 (32 of each class)	80.94%
10	MobileNet	32000 (synthesized data)	91.88%
10	MobileNet	41680 (full data)	95.63%

Sample Images Generated by the Conditional DCGAN

Acknowledgments

I would like to thank Soumith Chintala for his notes on training a GAN, which greatly helped me in developing the GAN architectures I used for this research.