Anthill Inside 2017

Availability of labelled data for supervised learning is a major problem for narrow AI in current day industry. In imaging, the task of semantic segmentation (pixel-level labelling) requires humans to provide strong pixel-level annotations for millions of images and is difficult when compared to the task of generating weak image-level labels. Unsupervised representation learning along with semi-supervised classification is essential when strong annotations are hard to come by.

This talk will introduce you to the techniques available in unsupervised learning and semi-supervised learning with specific focus on brain tumor segmentation from MRI using Stacked De-noising Auto-Encoders (SDAEs), which achieved competitive results in comparison with purely supervised Convolutional Neural Networks (CNNs), and highlight recent breakthroughs in AI for computer vision. Although the focus is on medical imaging, the techniques will be presented in a domain agnostic manner and can be easily translated for other sectors of deep learning.

Outline

Introduction - [10 mins]

• Deep learning in Medical Imaging
• Diagnosing Glioblastoma in brain with MRI
• Annotation problem
• Pre-processing

Auto-encoders - [15 mins]

• Auto-encoders
• Unsupervised learning
• Pre-training deep autoencoders on unlabelled data
• Fine-tuning on limited labelled data

Unsupervised learning - [5 mins]

• Novelty detection using autoencoder
• Segmentation using unsupervised learning

Results - [5 mins]

• Post-processing
• Segmentation results w.r.t state-of-the-art

Conclusions - [5 mins]

• Unsupervised + Supervised in one go
• Ladder Networks
• Y-Nets
• The future of unsupervised learning

Requirements

Fundamentals of supervised learning, convolutional neural networks, cost functions and over-fitting.

Speaker bio

Kiran Vaidhya holds a dual degree (B.Tech + M.Tech) in Engineering Design (specialization in Biomedical Design) from IIT Madras. He has been heavily involved in Computer Vision and Medical Imaging for the past 4 years. His Master’s thesis was on brain tumor segmentation from MRI using Semi-Supervised Deep Learning. His work has been published and accepted by leading medical imaging journals like MICCAI.

Post his graduation, he joined Predible Health and is currently working as an Algorithms Researcher for CAD (Computer Aided Diagnosis) system design. Deep learning is a natural part of his work in order to derive data-driven insights. He has been actively involved in the development of Torch and has extensive experience with Theano and TensorFlow.

Links

• Semi-supervised Learning using Denoising Autoencoders for Brain Lesion Detection and Segmentation - https://arxiv.org/pdf/1611.08664.pdf
• Multi-modal Brain Tumor Segmentation Using Stacked Denoising Autoencoders - https://link.springer.com/chapter/10.1007/978-3-319-30858-6_16
• LinkedIn - https://www.linkedin.com/in/kiran-vaidhya-2bb63aa8
• Predible Health - http://prediblehealth.com/
• GitHub - https://github.com/kvrd18

Slides

https://speakerdeck.com/kvrd18/unsupervised-and-semi-supervised-deep-learning-for-medical-imaging