<p>Outlier detection plays a critical role in various domains such as financial fraud detection, network intrusion detection, and medical anomaly analysis. Due to the widespread lack of labeled data in real-world scenarios, unsupervised outlier detection algorithms have emerged as a research hotspot. Among them, autoencoders have gained prominence as powerful deep modeling tools, primarily for their ability to quantify sample abnormality through reconstruction errors. However, training autoencoders on the entire dataset—including anomalous samples—often leads to impaired learning of the dominant data patterns, thereby degrading detection accuracy. To address this issue, this paper proposes an outlier detection algorithm that integrates topological modeling with selective sample filtering. The algorithm is built upon the collaboration between self-organizing maps (SOM) and autoencoders. Specifically, SOM is first employed to construct a low-dimensional topological representation of the input data. Guided by node density, a dynamic sample selection strategy is then applied to extract high-confidence potential normal samples from the raw dataset for training the autoencoder, effectively reducing the influence of anomalies during model optimization. Subsequently, the trained autoencoder collaborates by assessing the reconstruction error across the entire dataset to identify outliers with high precision. Extensive experiments conducted on 10 real-world datasets demonstrate the effectiveness and robustness of the proposed collaborative mechanism in comparison with several well-known algorithms.</p>