Computer Vision News - March 2020

2 Summary Code with Us 0 Deep learning research brings a lot of achievements in fields such as autonomous driving and importantly, in our case, the medical imaging field and its clinical applications. Unlike non-critical areas or toy applications where validation, explanation and accuracy aren’t so important, the clinical field is, as it is obvious, an area where even the tiniest error can have an irreversible or even lethal effect. A lot of discrepancies exist, as not everyone is ready to accept machine prognosis or even... diagnosis! Recently, a conversation started from a tweet by Geoffrey Hinton, about if a surgeon with 80% success rate should and could be compared with a “black box AI surgeon” with 90% success rate. The biggest fear of accepting deep learning in the clinical research is exactly that: AI is still a black box. Today, in this first part of the series, we will introduce how one can create a heart- condition model. In the second part, to be published on the April issue of this magazine, you will be able to see how to get insights from a deep learning model for a deeper understanding of its internal decision process. An example dataset of a heart disease prediction will be used (such datasets can be easily found online in popular websites like Kaggle or dataset search engines). Here is the code to get started: by Ioannis Valasakis A Step Towards Explainability in Deep Learning for Medicine Hi everyone! It is my pleasure writing for Computer Vision News, and I am honoured to take this journey with all of you. A short introduction about myself: I am Ioannis Valasakis, a software and hardware engineer with experience in the industry, various start-ups and research. I am currently pursuing a PhD in King’s London College, in the field of medical imaging.My research interest is inhowdeep learning techniques can be applied to cardiac image reconstruction. "The biggest fear of accepting deep learning in the clinical research is exactly that: AI is still a black box!"