Focal Adhesions Segmentation Using Deep Learning

Abstract

Focal Adhesions (FAs) are crucial for various cellular processes, including proliferation, survival, and migration. Traditional segmentation methods, such as the Focal Adhesion Cross-Correlation Kit (FACCK), often face challenges with complex FA structures and background noise, resulting in poor accuracy. This paper presents an advanced segmentation approach that employs U-Net and U-Net++ architectures, resulting in significant enhancements in accuracy. The proposed methods achieve Intersection over Union (IoU) scores of 72.8% and 73.85% for U-Net and U-Net++, respectively, showing a marked improvement over IOU FACCK's of 30.8%. Moreover, U-Net++ demonstrates considerable improvements in Precision (83.44%) and Recall (84.67%), surpassing FACCK's performance of 45.13% and 47.07%, respectively. These results highlight the potential of deep learning for FA segmentation, offering substantial benefits for biological research and enhancing our understanding of cellular behaviors