Deep AUC Maximization Method for Medical Image Analytics


In medical applications, an AI-based medical image classification system for identifying certain diseases serves a useful tool to aid medical diagnostic by doctors. It can also help to work with big data in a more efficient way.
The conventional solutions toward building such an AI-based classification system is to learn an artificial deep neural network on a labeled dataset by minimizing the error rate on the training data. However, these conventional approaches may suffer from low AUC score (area under the ROC curve) and are sensitive to outliers and had adverse effect when trained with well-classified data.
The researchers at the University of Iowa have invented a novel method to directly optimize AUC score for learning an artificial deep neural network; it consists of two steps: (i) pretraining a deep neural network with conventional approaches; (ii) fine tuning the deep neural network by maximizing a new objective function which serves as a surrogate for AUC.
The developed framework has also been carefully evaluated on two medical image classification competitions and achieved the state-of-the-art performance compared to existing solutions. The method can be used as a tool to automatically detect whether a medical image (e.g., Chest X-ray) indicates the patient has a certain disease.



Higher accuracy and lower error compared with conventional methods.

Contact Information:

Name: Kellen Sensor

Title :

Department :


Phone: 319.335.4546

Address :