This short course helps you improve segmentation models when classes are heavily imbalanced and predictions show recurring errors. You will learn how to apply class-balancing strategies such as focal-dice hybrid loss and sampling adjustments on medical or industrial datasets where foreground pixels may be extremely rare. You will also learn how to analyze predicted masks using region measurements to spot over-segmentation, under-segmentation, and shape-specific failures. Through concise videos, hands-on activities, and reflective checkpoints with Coach, you will practice improving recall, inspecting connected components, and building simple error logs that uncover patterns. By the end, you will have a repeatable approach for balancing datasets and diagnosing mask-level errors in production-ready segmentation workflows.