Data is crucial to build any deep learning models. Class imbalance is a known issue. For medical images, aim is to create class specific new image data using Generative Adversarial Networks (GANs) similar to Synthetic Minority Over-sampling Technique(SMOTE). The generated images can be used for data augmentation and to build a better classifier. GANs are a class of deep generative models which aim to learn a target distribution in an unsupervised fashion.
The speaker presents the details of Conditional Deep Convolution Generative Adversarial Networks (CDCGAN) and Auxiliary Classifier Generative Adversarial Networks (ACGAN). He discusses mode collapse problem and shares outline of the steps followed to make the GAN training stable. Initial results for medical images will be presented from the experiments carried out. The key takeaways are the identifying the challenges in generating class specific medical images using GANs and tricks to make training stable.
Dr. Sunil Kumar Vuppala works as a Principal Scientist and part of leadership team in Philips Research India, Bangalore. Sunil has 14+ years of industrial and research experience in Machine learning, Deep learning, Analytics, Optimization, Internet of Things, Automation, Healthcare Informatics and Smart Grid.
Before joining Philips, he worked in Infosys Labs of Infosys Limited, Bangalore for 10 years and in Oracle India Pvt. Limited, Hyderabad for 2 years. Sunil received B.Tech (CSE) from JNTU, Hyderabad in 2002 and M.Tech (IT) from IIT Roorkee in 2004 and Ph.D (IT) from IIIT Bangalore in 2018. Sunil is a lead inventor of 11 patents (6 granted and 5 pending applications). He has published more than 30 papers in both international journals and conferences and delivered 40+ guest lectures. He is a senior member of ACM and IEEE, fellow of IETE and life member of CSI.