Are ML Image Classifiers Robust to Medical Image Quality Degradation?

Abstract

Classification of medical images plays an important role in medical assisting applications, as it helps in performing routine patients diagnostics. In many cases, the images are transferred over a network to the cloud-based service that might result in their corruption. In this paper, we investigate how the ML medical image classification performance can be affected by the network Quality of Service (QoS) and Data Quality (DQ) degradation. In our study, we employ real-world X-ray image scans of lungs diseases and real life networks. As ML medical image classification systems, we employ well-known industrial VVG16, ResNet50, and InceptionV3 models, analyze and compare their performance. We leverage the POWDER platform to establish real wireless network between the two nodes. We transfer our X-ray scans between these nodes with various packet losses to obtain images corrupted due to network QoS degradation. We test our ML image classifiers on the obtained X-ray scans of various corruption degree and evaluate their performance. Our study demonstrates that even small packet losses of 2-5% can significantly deteriorate the ML classifier performance reducing the recognition accuracy from 90 to about 70-80 % that will make the cloud-based classification unacceptable for medical-related applications.