Student: Samuel Budd
1st supervisor: Bernhard Kainz, Imperial College London
2nd supervisor: Emma Robinson, King’s College London
Inaccurate data annotation, quality control and inaccessible data processing tools are a major problem in medical imaging research and diagnostics. Many modern medical image analysis methods that are based on machine learning rely on large amounts of annotations to properly cover the variability in the data (e.g. due to pose, presence of a pathology, etc). However, the effort for a single rater to annotate a large training set is often not feasible. To address this problem, recent studies employ forms of weak annotations (e.g. image-level tags, bounding boxes or scribbles) to reduce the annotation effort and aim to obtain comparably accurate results as to under full supervision (i.e. using pixelwise annotations) [Papandreou2015], [Dai2015], [Schlegl2015]. However, these tools are inaccessible to the wider research community and there is little emphasis on quality control.
In the biomedical visualisation and big data analysis domain, high performance computing is increasingly in demand for computations involving complex, resource-consuming models.
This project aims for a Service-Oriented Architecture (SOA) [Erl2005] that brings a compelling vision for medical image analysis. Web Services, as one of its popular instantiations, prompts open standards, loosely-coupled and platform-independent features, which can facilitate distributed computing, resource sharing, application interoperability and research collaboration. SOAs have not yet been considered for large scale medical image analysis as it would be required for exponentially increasing data set sizes.
Increasingly, large-scale, open, data collection projects are releasing large volumes of medical imaging data to the image analysis community. These include projects such as: BioBank, which is acquiring scans of all major organs for 100,000 volunteers between the ages from 40 to 69; the Human Connectome Project (HCP), which has released 1200 sets of high spatial and temporal resolution multi-modality Magnetic Resonance Imaging (MRI) data sets of adult human brains; and the Developing Human Connectome Project (dHCP), which stands to release 1500 data sets of functional, structural and diffusion imaging of fetuses, pre-term and term-born neonates. All data sets are motion compensated [Kainz2015] and supported by extensive databases of behavioural, genetic and demographic data providing a rich resource for machine learning.
The focus of this project will be the quality assured annotation of medical image data, initially human brain data using visual exploration in a SOA platform. Starting with adult human connectome data the project will eventually extend to the exploration of biomarkers and diagnosis of the developing human brain.
Data will be provided from the Developing Human Connectome Project – the first large-scale imaging data set to represent both healthy foetuses/term-born neonates and subjects born preterm, and those at risk of developing neurodevelopmental conditions. It presents a powerful resource from which biomarkers for developmental delay or conditions such as Autism may be identified.
The ultimate goal of this project will be to develop a web-application SOA framework to enable crowd-sourced segmentation [Rajchl2016] of fine scale structures in, for example, the fetal brain with automatic, machine learning-based quality control tools.
References:
[Papandreou2015] G. Papandreou, L.-C. Chen, K. Murphy, A. L. Yuille, “Weakly- and semi-supervised learning of a DCNN for semantic image segmentation” in , 2015, [online] Available: https://arxiv.org/abs/1502.02734.
[Dai2015] J. Dai, K. He, J. Sun, “Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation” in , 2015, [online] Available: https://arxiv.org/abs/1503.01640.
[Schlegl2015] T. Schlegl, S. M. Waldstein, W.-D. Vogl, U. Schmidt-Erfurth, G. Langs, “Predicting semantic descriptions from medical images with convolutional neural networks” in Information Processing in Medical Imaging, New York, NY, USA:Springer, pp. 437-448, 2015.
[Erl2005] T. Erl. Service-Oriented Architecture: Concepts, Technology, and Design. Prentice Hall, the United States of America, 2005
[Kainz2015] B. Kainz, et al.: Fast Reconstruction of Motion Corrupted Image Stacks with Automatic 3D Template Selection. IEEE Transactions on Medical Imaging, 34(9):1-13, Sept. 2015;
[Rajchl2016] M. Rajchl, M. C. H. Lee, O. Oktay, K. Kamnitsas, J. Passerat-Palmbach, W. Bai, M. Damodaram, M. A. Rutherford, J. V. Hajnal, B. Kainz, D. Rueckert, DeepCut: Object Segmentation From Bounding Box Annotations Using Convolutional Neural Networks, IEEE Transactions on Medical Imaging, volume 36, pp.674-683, 2017