I-Corps: Detecting Performance Degradation and Failures of Deep Neural Networks in Cancer Imaging

H. Lee Moffitt Cancer Center And Research Institute Hospital Inc, Tampa FL

Investigators

Abstract

The broader impact/commercial potential of this I-Corps project is the development of a failure detection framework that learns the behavior of the machine learning model under various noisy conditions. This solution is currently focused on cancer imaging applications, especially head and neck, lung and brain cancers. Neural networks are used in many areas of the human endeavor, and their use is expected to increase exponentially. These machine learning models do not provide a measure of confidence in the decisions and fail without warning. There are no solutions for addressing these issues except manually monitoring and reviewing the performance of these models after deployment. The proposed technology can integrate seamlessly with any machine learning model before or after deployment and output model confidence in the decision with minimal additional computational cost. The proposed technology will help artificial intelligence find its true potential in mission-critical areas. The proposed mechanisms for detecting performance degradation and model failure can provide a path to achieve the much-desired trustworthiness in artificial intelligence models. The applicability of the proposed technology encompasses various areas, including healthcare, transportation, cybersecurity, economics, environment, and financial services. This I-Corps project is based on the development of a generalized framework that quantifies the performance and detects failure in all types of machine learning models, including convolutional neural networks and transformers. This framework does not require retraining of the original model and can be used as an out-of-the-box solution. This technique consists of different methods to identify the type of machine learning model and its output. This information is used to specify a fixed threshold or learn a dynamic one. These threshold values serve as a guide for identifying the performance degradation of the machine learning model. In the first case, the technology defines a fixed threshold value based on the model performance on the test dataset with a changing signal-to-noise ratio. The second method learns the threshold value using a shallow neural network. The proposed failure detection methods seamlessly integrate with the original machine learning model and abstain from making decisions when the model’s confidence is below the threshold. This technique, when used during the machine learning model training phase, can help improve model accuracy. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.