<p>How computers can acquire the capability of radiologists to identify various anomalies within 3D medical images remains a central and challenging problem in intelligent imaging diagnosis. Supervised learning methods relying on lesion-level annotations are restricted to detecting a limited set of lesion types similar to those in the training data. Furthermore, obtaining voxel-level annotations is notoriously labor-intensive and time-consuming. To address these challenges, researchers have proposed unsupervised anomaly detection (UAD) methods for 3D medical images that learn from normal (healthy) samples. These approaches aim to identify any lesion that deviates from the learned normal anatomical distribution. Centered on this principle, three distinct UAD paradigms have emerged: self-supervised learning-based methods that train supervised models by constructing synthetic anomalies; deep feature embedding-based methods that measure the distance between test features and normal feature distribution; and reconstruction-based methods that transform abnormal images into pseudo-healthy images based on learned normal patterns. This paper provides a comprehensive review, analysis, and comparison of key studies within these three categories, highlighting their respective strengths, limitations, and potential research directions. Moreover, existing UAD methods are often validated using inconsistent datasets or preprocessing pipelines, making fair performance comparisons challenging. To address this issue, we establish a unified benchmark for evaluating UAD methods on 3D medical images, comprising nine brain MRI datasets (4 365 cases) and six liver CT datasets (376 cases). Using this benchmark, we evaluate the performance of 21 representative state-of-the-art algorithms on the 3D voxel-level anomaly localization task, objectively revealing the advantages and limitations of these methods.</p>