Egocentric hand-object interaction segmentation is a crucial task for understanding human behavior, yet existing methods face two major challenges: general segmentation models lack interaction relationship modeling, while specialized approaches overly rely on hand relationships and neglect detailed features of contacted objects, leading to poor segmentation of small objects and confusion between similar interactive objects. To address this, we propose FASAM2-Ego, a feature-enhanced SAM2 adaptation framework. Our main innovations include three aspects: First, leveraging DINOv2’s robust semantic features through a lightweight adapter to guide SAM2’s multi-scale feature fusion, enhancing perception of object details. Next, design a dual-decoder architecture where the mask generation branch inherits SAM2’s segmentation capability, while the semantic enhancement branch introduces directional attention for pixel-level semantic classification. Finally, we develop an end-to-end dual-encoder-dual-decoder model that integrates spatial-directional, channel, and pixel attention. Experiments on the EgoHOS dataset demonstrate that our method outperforms SOTA approaches in both in-domain and out-of-domain test.