A task-oriented review of deep learning for medical image segmentation from 2015 to 2026

Abstract

Segmentation of medical images is a fundamental step towards quantification and clinical decision support. Many approaches have been proposed and successfully demonstrated on benchmark datasets. However, reliable application of current approaches in real clinical practice remains challenging. Medical image segmentation is not a single task but rather a group of different tasks with different input and output characteristics, which call for dedicated solutions. Task difficulty and failure modes vary with anatomical priors, boundary ambiguity, and clinically meaningful error types. Therefore, rather than providing only a method summary, this review presents a task overview of medical image segmentation research from 2015 to 2026. We organize the literature according to clinical task properties and examine how different methodological choices influence benchmark results and real-world applicability. We categorize existing work into major task families, including organ segmentation, lesion and tumor segmentation, vascular segmentation, histopathology and cellular segmentation, and multi-task and cross-domain settings. For each category, we summarize representative modeling strategies and identify common challenges specific to each task. We further review widely used public datasets and evaluation metrics. In particular, we analyze how overlap, boundary, topology, and instance-level metrics capture different clinical priorities. Finally, we discuss challenges such as annotation burden, inter-institution variability, and clinical reliability. We then discuss emerging research directions, including foundation models, multi-modal and multi-task learning, and privacy-preserving federated collaboration, which may contribute to improved public health and clinical decision-making.