Image Downsizing and Compression Impacts on AI-based Medical Image Classification

Abstract

In recent years machine learning has been used for automatic medical image diagnosis. Solutions leveraging machine
learning can help physicians in the diagnosis process and also reduce the time spent by medical experts analyzing images and
video frames in order to conclude their assessments. In the application scenario considered in this work, the image is captured
in a resource-constrained remote basic health facility and sent to the inference model on the cloud that returns the result to the
physician. We investigate the impact of straightforward low-bit-rate image coding solutions on the classification performance of
neural network models targeted at cloud-based image diagnosis solutions. Our experiments show that it is possible to lower
the bit rate needs without significant harm to the prediction accuracy of the models using both downsizing and compression.
Thus, providing evidence for the viability of deploying automated diagnostic systems as a Service over constrained communication
infrastructure to assist remote areas.