Abstract: Interferometric Synthetic Aperture Radar (InSAR) provides wide-area, high-precision measurements of surface deformation and offers important data support for landslide detection. However, InSAR landslide targets often show large scale variation, fuzzy boundaries, and weak contrast with the background, which limit detection accuracy and localization quality. To address these problems, this study proposes an end-to-end InSAR landslide detection model named MSAM-DETR. Based on RT-DETR, the proposed model introduces shallow high-resolution features and Spatial-to-Depth Convolution (SPDConv) to preserve small landslide targets and boundary details. A CSP-OmniKernel multi-scale fusion structure is constructed and combined with the lightweight SimAM attention mechanism to enhance weak landslide responses and suppress background interference. In addition, an Inner-MPDIoU loss function is designed to improve boundary regression accuracy through an auxiliary-box mechanism and keypoint geometric constraints. Experiments on a publicly available InSAR landslide detection dataset show that MSAM-DETR achieves mAP@0.50 and mAP@0.50:0.95 values of 97.9% and 80.5%, respectively, outperforming RT-DETR by 2.9 and 8.0 percentage points. Ablation experiments further demonstrate the effectiveness of the proposed modules in multi-scale feature representation, weak-contrast landslide detection, and boundary localization. The results indicate that MSAM-DETR improves high-threshold localization quality while maintaining high detection accuracy and inference efficiency, providing technical support for wide-area landslide surveys and early identification.