Medical Image Data Cleansing for Machine Learning: A Must in the Evidence-Based Medicine?
摘要
Preparing the data for machine learning is important, as it has been proven that the quantity and quality of the input data is a strong predictor for the output. In a deep-learning with transfer learning context, an ultrasonography image dataset is cleaned. The performance of the cleaned dataset against the original one is measured using the accuracy and area under the curve metrics running three different network architectures: AlexNet, GoogLeNet, ResNet-18. Both metrics show significantly superior results on the cleaned dataset, though 49.81% of the original dataset has been removed. We conclude that data cleansing must be done before applying any machine learning algorithms, physicians should cope with the machine learning terminology and focus on understanding the models in use, and last but not least, the data cleansing should be driven by computer scientist though they have less understanding of the medical problem itself.