Segmenting Fluorescent-Dyed Cell Images Using Mask R-CNN

Abstract

This project explores the impact of contrast-related hyperparameters on cell detection performance in fluorescent-labeled images. We focused on optimizing model architectures and training strategies, with an emphasis on static and dynamic auto-contrast methods. Our findings show that SGD with momentum consistently outperformed other optimizers across dynamic auto-contrast configurations. Static auto-contrast methods, particularly those using Pillow and PyTorch, enhanced edge detection, but occasionally hindered the visibility of fluorescent Kog1 bodies crucial for classification. Kernel size variations in the Basic Convolutional Block module did not significantly affect performance. Data augmentation proved effective in improving model generalization on a small dataset. Among the dynamic auto-contrast models, the Baseline Mask R-CNN model provided the best performance. In support of our hypothesis, static auto-contrast methods improved segmentation accuracy by enhancing edge definition and feature visibility. Our best-performing model achieved segmentation accuracy within 5% of ACCT, a state-of-the-art model for brightfield images, demonstrating the value of contrast-aware preprocessing in fluorescent image segmentation.