Artificial Intelligence-Based Diffraction Analysis (AIDA) for Point-of-Care Breast Cancer Classification
Abstract
Background: A disproportionate number of minority women present with aggressive breast cancers due to biological and access reasons. While microscopes remain integral tools for diagnosing cancers and informing treatment decisions, their low throughput and need for highly trained personnel can prove limiting. These drawbacks are more acutely experienced in resource-limited settings (e.g., rural areas, developing countries, remote military installations) where severe pathology bottlenecks delay cancer diagnoses and lead to over-/under-treatment. Improved diagnostic capacities to feasibly detect and classify breast cancers at the point of care (POC) is thus a key mandate for better cancer management and improved survivorship. Unfortunately, no such platforms are currently available for translational testing. Overarching Challenges: This proposal will address the following challenge area: • Distinguish deadly from non-deadly breast cancers. • Conquer the problems of overdiagnosis and overtreatment. Objective: Develop a high-throughput, intelligent diagnostic platform – AIDA. We will advance a novel diagnostic technology, AIDA (Artificial Intelligence-based Diffraction Analysis), for automated and highly reliable breast cancer cell screening. The AIDA will integrate two cutting-edge developments in the computer age: computational microscopic imaging and deep neural networks. Digital diffraction imaging is a novel approach that offers wide field-of-views in contrast to standard microscopes. The method records the inherently complex patterns generated by shining light on microscopic objects (diffraction) and then digitally restores the patterns into the original images. The system is very simple to build, not requiring additional optical components, and thereby easy to operate and cost-effective. Deep neural networks (DNN) depict the process of training computers to learn from data sets, eventually figuring out its own solutions to given problems. DNN is making impressive progress in its ability to derive information from big data. It is a highly potent approach for discovering intricate, hidden structures within very complex data sets. This proposal will exploit the power of DNN to analyze raw diffraction images. We specifically hypothesize that deep learning can reliably recognize breast cancer cells and extract any needed molecular information directly from inherently complex diffraction patterns. By combining digital diffraction with DNN, we expect to create a robust, portable, and user-friendly system for large-scale imaging. Specific Aims: Aim 1: Implement an AIDA platform for large-scale single cell imaging. Aim 2: Develop AIDA deep-learning framework for single cell detection and classification. Aim 3: Apply AIDA to profile clinical breast cancer cell samples. Study Design: Aim 1: We will develop AIDA imaging hardware with multiple light sources to generate unique light beams. Such modifications will allow us to detect different cancer markers on cells through simultaneous color-based testing. We will also establish and optimize special protocols to enable simultaneous analyses of breast cancer markers on individual cells. Aim 2: We will establish and optimize special protocols to enable simultaneous analyses of breast cancer markers on individual cells. Specifically, we will train a DNN to (i) recognize individual cells directly from raw diffraction images and (ii) extract various levels of biologically relevant information. The trained network and the imaging system will then be applied to analyze breast cancer cells. We will focus on detecting markers (i.e., HER2, ER, PR) often tested per clinical routine. Aim 3: We will perform a clinical study to test AIDA’s clinical utility. Samples will be obtained from breast cancer patients and molecularly profiled by applying AIDA test results will be compared with current standard methods (flow cytometry, histology). Impact
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 16, 2019
- Source ID
- W81XWH1910199
Entities
People
- Hakho Lee
Organizations
- Massachusetts General Hospital
- United States Army