We propose a study of radiomic texture analysis in terms of robustness assessment and classification utility. We will introduce novel robustness metrics geared towards assessment of radiomic features in comparison across two image conditions, and apply these metrics to study feature robustness across imaging parameters and patient biology. In addressing the utility of radiomic features in cancer risk assessment, we will identify and evaluate texture signatures from mammography and tomosynthesis datasets. The field of radiomics is evolving fast, and quantitative texture analysis is being applied to a growing number of applications in medical imaging. By performing a thorough investigation of the robustness of these radiomic features to dataset heterogeneities we aim to identify the strengths and weaknesses of commonly used features to guide their implementations on future applications. Two clinical tasks will be studied under the proposed research: 1) risk assessment and cancer prediction and 2) malignancy evaluation. Multiple modalities including tomosynthesis, mammography and MRI will be involved in studies geared towards addressing these clinical questions. An evaluation of the robustness of commonly employed radiomic features will help guide the field of medical texture analysis and contribute to meaningful conclusions in future studies throughout the field of quantitative image analysis.
The first aim of the proposed research involves the proposition and evaluation of novel robustness metrics for investigations lacking a classification task.
The second aim will extend the study of radiomics to investigate the utility of robust features in classification tasks and identification of texture signatures relate to biomedical characteristics.
The third aim will build upon the two previous aims and culminate in the application of cutting-edge technologies in machine learning and deep learning in further promoting image processing in the field of medical physics.
The goal of the proposed research is to evaluate and improve the application of radiomic texture features in cancer risk assessment. We will accomplish this by evaluating the robustness of various radiomic metrics, testing the classification utility of texture features in clinical tasks, and extending current classification methods to include cutting-edge developments in machine learning technology. Careful preliminary studies have demonstrated methods for selection of robust texture features and improvement in classification tasks by emphasizing feature robustness in feature selection methodology and we therefore believe that a meticulous evaluation of the impact of imaging parameters on feature calculations will lead to overall improvement of computer-aided diagnosis and clinical translation to progress in cancer screening protocols.
