From jets to MRI machines, flawless metal components are vital across industries. But defects sneak past human inspectors reviewing electron microscope images. We will demonstrate AI that rapidly segments defects at 10x human accuracy. Our neural network architecture performs semantic segmentation, intelligently tracing helium bubbles in microscope images. Unlike previous models, we employ novel distance-based and dice loss functions to delineate precise defect boundaries for tracking material swelling. Careful loss function selection proves critical. Standard approaches struggle with imbalanced datasets and our metal alloy defects. However, we will show that combining distance map and dice losses significantly improves the AI model. Our enhancements produce superior defect segmentation, advancing reliability assessments. This project exemplifies the importance of tailored model selection, customized objective functions, and multifaceted performance metrics. With thoughtful tuning, we demonstrate reliable automation of tedious microscopy workflows. Our techniques can generalize across sectors requiring meticulous material screening.