Developing Machine Learning Tools for Quantitative Analyses of Biomedical Images
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Biomedical and pathology images have a significant role in medical and research applications in both diagnosis and disease treatment. With the evolution of digital scanning tools, increasing computing power, and deep learning tools, it is necessary to provide researchers with accurate, fast, and user-friendly tools that can be used in clinical and research applications. In this dissertation, we applied these tools to two biomedical areas and diseases: the first is obesity by studying adipose tissue histology images; the second is breast cancer by studying tumor histology slides. First adipose tissue images were analyzed using the AdipoGauge software that we developed. This software can detect adipose cells, count them, and calculate the area of all cells very accurately. It also can find areas of interest in adipose tissue and calculate the required data for research purposes such as cell size and cell number in different biomedical slides. AdipoGauge contains different analysis tools such as object removal, bordering cell detection, and cell size categorization which makes the analysis process more accurate. We demonstrated that the results from the AdipoGauge software were more accurate than similar software such as ImageJ. Next, the framework of the AdipoGauge was used to develop a new tool for quantitative analyses of breast tumor histology images. Detection of the Tumor Infiltrating Lymphocytes (TILs) in cancer images has gained significant importance as it can be used as a biomarker to guide cancer detection and treatment. Based on the guidelines from the International Immuno-Oncology Biomarker Working Group (IIOBWG) on Breast Cancer, we labeled 63 large pathology imaging slides and annotated the TILs in the stroma area to create the dataset required for model development. A U-Net deep learning model was employed and trained for different parameters to achieve the highest accuracy possible which is 98% accuracy in the stroma detection task. The segmented stroma images then were used to implement the Mask R-CNN model for the TILs detection task. The R-CNN model detected the TIL cells in different images and was used as the backbone analysis network for the GUI development of the TILs detection tool. In conclusion, we developed sophisticated machine learning tools and software that will assist researchers and clinicians in rapid and accurate analysis of histology and pathology slides, which will help future research and/or treatment strategies.
Embargo status: Restricted until 06/2028. To request the author grant access, click on the PDF link to the left.