ICL-NoiseUNet - A Novel In-Context Learning Based Framework For Ultrasound Segmentation With Adaptive Noise Modulation

The complex patterns, artifacts and speckle noise that are present in ultrasound images make precise segmentation very challenging. Existing approaches, such as convolutional neural network architectures and foundation models, have shown promising results across a wide range of tasks. However, they struggle to adapt to the unique characteristics of ultrasound data, leading to poor delineation of anatomical boundaries. For that reason, we propose ICL-NoiseUNet, an in-context-learning segmentation framework that combines guidance from a set of input-output pairs, called the context set, with analytic noise descriptors. More specifically, the model leverages an In-Context Feature Conditioning (ICFC) module to incorporate context examples and a Noise Modulation Block (NMB) that adapts feature representation to ultrasound characteristics. After extensive evaluation across several datasets, ICL-NoiseUNet consistently outperforms state-of-the-art methods, enhancing the segmentation quality. Moreover, ablation studies confirm the synergy effect of contextual conditioning and noise modulation. Overall, these findings pave the way for noise-guided ultrasound segmentation. The code will be open-source at .