Project Summary Lack of transparency and trustworthiness of deep neural networks (DNNs) has long been recognized as a major drawback of the technology, hindering its widespread acceptance in many practical applications. The objective of this project is to establish a novel contrastive feature analysis (CFA) framework for reliable visualization of the high dimensional feature space and effective design of high-performance DNNs for medical image analysis. We hypothesize that CFA-based feature visualization will enable us to quantify the quality of the feature space at different layers during training/testing of a DNN and empower us with an effective tool to prune the network architecture for enhanced performance. Specifically, we will (1) develop an efficient visualization technique CFA for high dimensional feature data, 2) apply the CFA visualization framework to automatically refine DNN architecture for improved performance, and 3) demonstrate the potential of CFA in solving clinical problems. Successful completion of the project will enable us to analyze the feature data reliably and quantify the quality of the feature space at different layers of a DNN. The study also promises to provide high-performance DNNs for medical image analysis to substantially improve the AI-based diagnosis, prognosis and treatment planning of different diseases.

Project Narrative The proposed research is directed at establishing a novel strategy of analyzing the feature space data extracted from medial images by deep neural networks (DNNs). Successful completion of the project will enable us to analyze the feature data reliably and quantify the quality of the feature space at different layers of a DNN. The study also promises to provide high-performance DNNs for medical image analysis to substantially improve the AI-based diagnosis, prognosis, and treatment of different diseases.