SAM+D: Parameter-Efficient Dimensional Lifting of SAM-Family Models via Depth-Routed LoRA and Depth Shifting

Researchers introduced SAM+D, a parameter-efficient fine-tuning framework that adapts frozen SAM and SAM2 models to 3D and 4D segmentation.

It uses depth-routed LoRA experts and depth shift modules to share information across slices or time while keeping most pretrained weights fixed.

On 3D CT benchmarks such as KiTS, Pancreas, LiTS, and Colon, as well as 4D datasets like Fluo-N3DH-SIM+ and CTC, it matched or beat prior methods with single-point prompts.

The work shows large segmentation models can be extended to volumetric and spatiotemporal data without major architectural changes or full retraining, cutting compute and parameter costs.