Maggnus - Yale Hack Health 2018

Visualization of ultrasound classifier

The challenge was to interpret the performance of inception v1 network on their ultrasound images gathered from using Butterfly hand-held ultrasound devices. We utilized a simple method of erasing parts of images, feeding them to the classifier, observing the class probability of the correct class. White means higher value of class probability, meaning the model was more sure of its prediction when that particular region was removed. Dark means low value of class probability, meaning that dark regions are the important regions for the model to make correct predictions.

Jungkyu Park
Deep Learning Researcher, PhD Student

Efficient Deep learning in medical imaging