Segment Anything Model (SAM) has achieved significant success in natural image segmentation. However, it struggles with ultrasound images, especially when dealing with individual organs with small samples. To address these challenges, we propose SAMTNU, an adaptive Segment Anything Model for Thyroid and Nodule Ultrasound Image Segmentation. The SAMTNU consists of two primary branches: The SAM branch and the Convolutional Neural Network (CNN) branch. Specifically, we fine-tune the SAM branch’s image encoder using adapters and a Multi-head Attention Low-Rank Adaption module (MHA-LoRA) while keeping other parameters frozen, reducing training costs. To extract more comprehensive features, we propose a Multi-Scale Attention Feature Extraction module (MSAFE) in the CNN branch. Experiments on three datasets show that SAMTNU outperforms existing promising methods. For example, SAMTNU achieves Dice \(\uparrow \) of 83.87% and HD \(\downarrow \) 23.98 on TG3K, surpassing the SAMUS by 7.49% and 2.56, respectively.