2019 — 2021 |
Harkness, John H |
R43Activity Code Description: To support projects, limited in time and amount, to establish the technical merit and feasibility of R&D ideas which may ultimately lead to a commercial product(s) or service(s). R44Activity Code Description: To support in - depth development of R&D ideas whose feasibility has been established in Phase I and which are likely to result in commercial products or services. SBIR Phase II are considered 'Fast-Track' and do not require National Council Review. |
Development of An Adaptive Machine Learning Platform For Automated Analysis of Biomarkers in Biomedical Images @ Rewire Neuroscience, Llc
ABSTRACT Manual analysis of biomedical images by researchers and pathologists has the potential to introduce bias and error that compromise the reliability of research and clinical findings. These problems are significant barriers to delivering the most beneficial evidence-based medicine, developing effective medical treatments, and promoting confidence in scientific inquiry. Identification of biomarkers and cellular targets following microscopy requires manual analysis of biomedical images, which is time intensive, difficult, and prone to bias and errors. Unintentional bias and attentional limitations during analysis of biomarkers can underlie poor reproducibility of findings in biomedical research and potentially introduce error in clinical diagnostics. We recently developed a ?beta? software package designed to improve automation and standardization of image analysis, called ?PIPSQUEAK? (Perineuronal net Intensity Program for the Standardization and Quantification of Extracellular matrix Analysis Kit). Since its publication in 2016, PIPSQUEAK beta has amassed approximately 1,300 users worldwide who use it to quantify the intensity and number of perineuronal nets and other neural markers in the brain. This technology significantly increases data reliability between image raters and decreases the time required for analysis by more than 100-fold. However, PIPSQUEAK beta currently uses target detection algorithms that require high-contrast images to automatically identify neurons as clusters of bright pixels on dark backgrounds. A significant current limitation to PIPSQUEAK beta, and other available imaging programs, is that detection of biomarkers can be difficult unless image conditions are ideal. Suboptimal conditions, like high background staining, off-target structures, overlapping or clustered biomarkers, and atypical morphologies, can lead to artifacts and consequently to inaccurate results and erroneous conclusions. Here, we propose to develop a user-friendly artificial intelligence (AI) platform for the automated detection of targeted biomarkers in digital microscopy that reduces this error by learning to distinguish between true cellular biomarkers and artifacts. We propose to integrate AI capabilities into our PIPSQUEAK technology to produce an adaptive, high-throughput, biomedical image analysis platform that quickly and accurately identifies biomarker targets from bench to bedside. A key advantage is that this AI program will be user friendly and available online, making it highly accessible to basic researchers and to technicians and clinicians identifying human pathologies. Thus, successful development of our AI program has a high translational potential. The goal of this proposal is 1) to develop and validate a machine learning model that is capable of detecting common histological marker morphologies in digital microscopy, and 2) to test the feasibility of adapting our AI platform to new biomarker datasets with minimal additional supervised training. Our end goal is to advance the reliability and speed of research findings and clinical diagnoses by making this technology widely available to researchers and clinicians.
|
0.921 |