<p>Instance segmentation, the pinnacle of image analysis tasks, synergizes object detection’s localization prowess with semantic segmentation’s pixel-level precision. Over the past decade, the field has experienced a profound transformation driven by deep convolutional neural networks, large-scale annotated datasets, and, most recently, vision transformers and foundation models. This systematic review synthesises research papers spanning 2017 to 2025 to provide a coherent, beginner-friendly pathway from classical image processing roots to the latest developments in real-time, 3D, and open-vocabulary instance segmentation. We establish a four-class taxonomy—top-down (detect then segment), bottom-up, one-stage, and transformer-based methods, and trace how each paradigm emerged in response to limitations of its predecessor. Key milestones discussed include the region proposal framework introduced in R-CNN, the landmark Mask R-CNN architecture, the real-time YOLACT and YOLACT++ systems, and modern query-based transformers such as FastInst. We further survey a broad spectrum of application domains: medical imaging (brain tumors, liver, lung nodules, nuclei), remote sensing (SAR ships, buildings, UAVs), agriculture (crops, livestock, aquaculture), and emerging areas such as audio-visual, 4D, and open-vocabulary segmentation. Evaluation metrics, benchmark datasets including Microsoft COCO, PASCAL VOC, and domain-specific corpora, and the persistent challenges of small objects, occlusion, real-time inference, and data scarcity are critically examined. The review concludes with a forward-looking discussion of diffusion-model-based segmentation, segment-anything paradigms, and multimodal instance understanding as the next frontiers of the field.</p>