- Qure.ai has developed machine learning algorithms to detect abnormalities in head CT scans
- The researchers trained the model on a dataset of 310k images, out of which 21k images were held out for the validation set
- The final model revealed an accuracy of 95% on the validation set
Machine learning has made a significant impact in healthcare but the general perception is that there is still a long way to go before we can fully trust machines to make life and death decisions. It’s an understandable line of thought, but a few researchers are breaking new ground in their attempts to improve on the existing healthcare technology.
One of them is Qure.ai, a company aiming to revolutionize healthcare with the power and assistance of deep learning. When a patient comes in with a head injury, time is of the utmost importance. A few seconds here or there can make all the difference. Qure.ai has developed their algorithms such that they can analyse CT scans of the brain in less than 10 seconds.
But these algorithms, as you can imagine, require a lot of data to be built upon. The research team at Qure.ai collected a dataset of 313,318 images of head CT scans plus their medical reports from multiple centres. Out of this data, 21,095 scans were held out to validate the final model. The remaining data points were used to develop the algorithms. The final model has shown a 95% accuracy on the validation set.
Qure.ai can also develop automated reports. Their algorithms can localize and quantify haemorrhages and fractures. Put together with brain anatomy segmentation algorithms, their machine learning powered system can automatically generate reports, an example of which you can see below:
The findings from this study were published in a research paper called “Development and Validation of Deep Learning Algorithms for Detection of Critical Findings in Head CT Scans”. You can also check out their GitHub library here to follow their progress.
Our take on this
We have recently seen Google’s heart disease and cancer detection algorithm and IBM’s psychosis detection algorithms. Qure.ai has put forward a worthy addition to that list. This is quite a useful tool for radiologists. With the high level of true positives, this can certainly be used to assist decisions (rather than replace doctors and assistants, as is the general fear). Anyone in the field of deep learning, or interested in this field, should download their dataset and work on it. It will not only enhance your data science skills but might end up helping the world of healthcare, and consequently, the community as a whole.
I highly recommend reading their research paper and then trying out their techniques on this dataset. Let us know your findings and thoughts in the comments section below!
Subscribe to AVBytes here to get regular data science, machine learning and AI updates in your inbox!
You can also read this article on our Mobile APP