Medical imaging and diagnostics
It is estimated that one in six people will suffer a stroke in their lifetime. Six of the estimated fifteen million global victims die every year and another six million incur a permanent disability. To minimise the immense burden of this devastating disease, physicians are faced with the challenging task of quickly evaluating brain scans to initiate treatment as soon as possible.
A general lack of available CT scan data is limiting the performance of deep learning models. Without sufficient data to learn from, neural networks struggle to generalise and learn patterns.
We’re exploring cutting edge techniques to discover workarounds and compensate for the missing annotated data. We’re testing four brand new methods around convolutional neural networks.
- Weak labelling: Training neural networks with approximate labels instead of absolute labels.
- Redimensioning: Learning from a smaller set of larger images by applying ‘sparse convolutional kernels’ and ‘octree-based CNNs’.
- Heterogenous data: Training CNN’s on data from different types of scanners to make it more robust.
- Capsule networks: Adding structures called capsules to a CNN and reusing output from several of those capsules to form more stable representations for higher order capsules.
- Break the 80% accuracy barrier by more effective usage of available data.
- Enable fast and accurate image analysis that helps physicians in making well-informed decisions.