Unsupervised learning from massive scale medical image collections
Background: UK Biobank is the world’s largest population health study, collecting health and lifestyle records from over 500,000 subjects in the UK. For a subset of 100,000 subjects UK Biobank is also collecting medical images in form of Magnetic Resonance Images (MRI) of the brain, heart and whole-body. However, only a small number of the images have pixel- or voxel-label labels which can be used for supervised learning approaches.
Challenge: The aim of the project is investigate unsupervised learning approaches for large collections of medical images. The project has two aims:
- To use simple image-level tags (image sequence, image plane, organs) to learn a powerful CNN-based model (similar to ImageNet) that can be refined using a small amounts of pixel-level labels to perform tasks such as semantic segmentation.
- To experiment with self-supervised learning techniques such as described in these papers (paper 1, paper 2) in order to build powerful CNN-based models.
Skills: The project will require a good understanding of machine learning techniques as well as some knowledge of computer vision. The project also requires very good programming skills.
Using deep learning to generate semantic descriptions for images
Challenge: This project will explore the use of deep learning techniques such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to generate natural language descriptions of images and objects within these images. The aim to develop and implement technique that can learn the inter-modal correspondences between language and visual representations.
There are several possible approaches to implement this. The first approach aims to align both modalities (text and images) in a multimodal embedding. This inferred alignment is then used to learn to generate novel descriptions of image regions. This approach is described in the following paper: Deep Visual-Semantic Alignments for Generating Image Descriptions
A different approach is to generative model based on a deep recurrent architecture that combines recent advances in computer vision and machine translation and that can be used to generate natural sentences describing an image. Such as an approach is implemented in the following paper: Show and Tell: A Neural Image Caption Generator
The project will develop similar approaches to those described above and evaluate how well these perform in data from real-world settings.
Skills: The project will require a good understanding of machine learning techniques as well as some knowledge of computer vision. The project also requires very good programming skills.
Visualization of massive-scale medical image datasets using the DSI Data Observatory
Background: UK Biobank is the world’s largest population health study, collecting health and lifestyle records from over 500,000 subjects in the UK. For a subset of 100,000 subjects UK Biobank is also collecting medical images in form of Magnetic Resonance Images (MRI) of the brain, heart and whole-body. State-of-the-art machine learning techniques for dimensionality reduction such as manifold learning provide powerful approaches to uncover the relationship between images (i.e. their similarities and dissimilarities). However, the intuitive visualization of these relationships is still challenging.
Challenge: The aim of this project is three-fold:
- To implement state-of-the-art dimensionality reduction techniques for massive-scale medical image datasets such as UK Biobank.
- To explore visualization techniques for massive-scale medical image datasets such as UK Biobank to allow the user to interactively query the image databases
- To implement a web-based user interface for visualization using the facilities of the visualization studio of the Data Science Institute at Imperial.
Skills: The project will require a good understanding of machine learning techniques as well as some knowledge of computer graphics and/or computer vision. The project also requires very good programming skills.