We propose to introduce and optimize a new method of radiomics extraction via transfer learning with deep convolutional neural networks (CNNs) and to compare it to the conventional segmentation-based radiomics approach on breast dynamic contrast-enhanced magnetic resonance images (DCE-MRIs). The field of breast radiomics has been expanding fast, with many clinical conclusions being successfully derived from medical images using qualitative analysis. In the past couple of years, deep learning has experienced explosive growth in image recognition, easily solving complex problems. Deep CNNs achieve remarkable classification results on everyday image datasets. We propose to investigate the utility of deep neural networks with regards to the medical image datasets, specifically on the breast DCE-MRI dataset. Given the relatively small sizes of these datasets, CNNs previously trained on non-medical images will be utilized for clinical classifications as feature extractors. We will investigate multiple parameters involved in the CNN feature extraction methodology and their effect on classification performance. Two clinical tasks will be studied under the proposed research: 1) malignancy assessment and 2) response to therapy prediction. The optimized CNN method will be compared to and combined with the conventional segmentation- based radiomics method. Furthermore, we aim to investigate the robustness of the segmentation-based features across MR scanners of different manufacturers.
The first aim of the proposed research will study the robustness of the segmentation- based features extracted from images acquired on MR scanners of two different manufacturers. The robustness will be investigated under four clinical tasks, such as lymph node involvement and receptor statuses.
The second aim will be focused on optimization of CNN feature extraction and subsequent classifier design. Lastly, under the third aim we will compare and combine the CNN and segmentation-based radiomics in the classification tasks of malignancy assessment and response to therapy prediction.

Public Health Relevance

The goal of the proposed research is to improve breast cancer diagnosis and prognosis based on dynamic contrast-enhanced magnetic resonance images by introducing novel deep learning methods to medical image classification and combining it with the conventional radiomics systems. The incredible power of deep learning methods to classify everyday images shows great promise to make predictions based on medical image datasets. Our thorough investigation of deep learning methods and their combination with conventional radiomics methods has potential to improve breast cancer management.

Agency
National Institute of Health (NIH)
Institute
National Cancer Institute (NCI)
Type
Predoctoral Individual National Research Service Award (F31)
Project #
1F31CA221193-01A1
Application #
9469826
Study Section
Special Emphasis Panel (ZRG1)
Program Officer
Radaev, Sergei
Project Start
2018-03-01
Project End
2021-02-28
Budget Start
2018-03-01
Budget End
2019-02-28
Support Year
1
Fiscal Year
2018
Total Cost
Indirect Cost
Name
University of Chicago
Department
Type
Schools of Law or Criminology
DUNS #
005421136
City
Chicago
State
IL
Country
United States
Zip Code
60637