This is a follow-up study for our funded NIH R21 project (EB019074, 09/11/2015-06/30/2019) and the previous non-disclosure agreement (NLST-53). it is clear that the current best strategy is comprehensive diagnostic testing and adaptive individualized therapy. Therefore, research on sophisticated biomarkers is key, and imaging informatics must perform exclusive and intelligent mining through a huge amount tomographic data to help build correlative and predictive models. The overall goal of this project is to develop a tensor-based deformable dictionary learning (DcL) method for extraction of CT imaging biomarkers and several deep learning (DpL) algorithms to compress these biomarkers for differentiation between true-false and positive-negative CT lung screening results. The major innovation of this project is to synergistically integrate tensor decomposition, deformable dictionary learning, compressive sensing, greedy layer-wise training of deep networks, deep Boltzmann machine, multi-class boosting, supercomputing and big data mining into a brand-new imaging informatics approach. The three specific aims are defined as follows.
Specific Aim 1 – Deformable Dictionary Learning-based Image Feature Extraction
Task 1.1. Dictionary Learning Framework: State-of-the-art tensor decomposition and deformable dictionary techniques will be used to formulate a discriminative DcL framework for lung cancer screening. The DcL-based sparse representation, in terms of cancer-relevant dictionary atoms, and their relationships will be treated as novel phenotypic biomarkers.
Task 1.2. Functional Optimization: A block successive upper-bound minimization scheme will be used to optimize the deformable DcL objective functional.
Specific Aim 2 – Deep Learning-based Feature Extraction and Classification
Task 2.1. Discriminative Feature Learning: Based on recent advances in DpL, a new multimodal learning method based on greedy layer-wise training of deep networks and deep Boltzmann machine learning method will be developed to extract discriminative features.
Task 2.2. Classification and Multi-class Boosting: A two-level classification scheme and multi-class boosting classification approach will be developed to classify the lung cancer screening results into negative (normal) and positive (lung cancer), as well as to stage positive results. Visualization techniques will be developed for post-hoc interpretations.
Specific Aim 3 – Clinical Application and Evaluation
Task 3.1. Software Development: A user-friendly software interface will be developed. Then, all the components of the proposed algorithms will be integrated and verified using 200 selected cases from the NLST database.
Task 3.2. Computational Acceleration: The proposed algorithms will be accelerated via different software and hardware methods to meet research and clinical needs, as outlined in Task 3.3.
Task 3.3. Observer Study: A pilot study will be performed to show that the proposed imaging informatics methodology can help to reduce the false-positive rate while the true-positive rate is maintained. An observer study will be performed to evaluate the ability of the output generated by the proposed imaging informatics methodology to assist radiologists in the distinction between benign and malignant lung nodules. The performance of the radiologists first without, and then with, the computer output will be evaluated through receiver operating characteristic analysis.

Yu Cao, PhD, Professor, University of Massachusetts Lowell