The Segment Anything Model and its variants, such as MedSAM, have demonstrated potential for medical image segmentation. However, they heavily rely on high-quality manual prompts, which are both time-consuming and require domain expertise. Even when using manual prompts, including sparse prompts like boxes, points, or text, and dense prompts such as masks, SAM and its variants like MedSAM (fine-tuned on medical images) fail to segment bones in ultrasound images due to significant domain shift. To address these limitations, we propose Sam2 for Radiology (Sam2Rad), a framework that extends SAM and its recent iteration SAM2 to segment bony regions in ultrasound images without requiring manual prompts. At the core of our approach is a Prompt Predictor Network (PPN) that uses a lightweight cross-attention mechanism to generate bounding box coordinates, mask prompts, and high-dimensional embeddings to be used as prompts. Specifically, PPN leverages hierarchical feature maps extracted from SAM's image encoder as keys and values, and learnable object embeddings as queries. The output of the cross-attention is then used to predict the bounding box coordinates, mask prompts, and high-dimensional prompts. These predicted prompts are subsequently fed to SAM's mask decoder to generate the final segmentation mask. By aligning the learned prompts with SAM's original training scheme, PPN enhances compatibility with SAM's architecture, requiring no additional standalone encoders. To preserve SAM's extensive world knowledge, we keep all SAM modules frozen and train PPN only. This approach enables efficient parameter utilization while retaining SAM's generalization capabilities. Additionally, Sam2Rad can operate in three modes: fully autonomous without human supervision, semi-autonomous with human-in-the-loop for iterative refinement, and fully manual for tasks like data labeling. We tested the proposed model - Sam2Rad on 3 musculoskeletal US datasets - wrist (3822 images), shoulder rotator cuff (1605 images), and hip (4849 images). Without Sam2Rad, all SAM2 variants failed to segment shoulder US in zero-shot generalization with bounding box prompts. Our model, Sam2Rad, improved the performance of all SAM base networks in all datasets, without requiring manual prompts. The improvement in dice score ranged from a 2.2%-5.8% for hip, 19.6%-32.8% for wrist wrist, and up to 51.3% improvement in Dice score (from 25.2% to 76.5% Sam2 large) on shoulder data. Notably, Sam2Rad could be trained with as few as 10 labeled images and it is compatible with any SAM architecture. The code is available at https://github.com/aswahd/SamRadiology.

Lung ultrasound (LUS) is a non-invasive bedside imaging technique for diagnosing pulmonary conditions, especially in critical care settings. A-lines and B-lines are important features in LUS images that help to assess lung health and identify changes in lung tissue. However, accurately detecting and segmenting these lines remains challenging, due to their subtle blurred boundaries. To address this, we propose TransBound-UNet, a novel segmentation model that integrates a transformer-based encoder with boundary-aware Dice loss to enhance medical image segmentation. This loss function incorporates boundary-specific penalties into a hybrid Dice-BCE formulation, allowing for more accurate segmentation of critical structures. The proposed framework was tested on a dataset of 4599 LUS images. The model achieved a Dice Score of 0.80, outperforming state-of-the-art segmentation networks. Additionally, it demonstrated superior performance in Specificity (0.97) and Precision (0.85), with a significantly reduced Hausdorff Distance of 15.13, indicating improved boundary delineation and overall segmentation quality. Post-processing techniques were applied to automatically detect and count A-lines and B-lines, demonstrating the potential of the segmented outputs in diagnostic workflows. This framework provides an efficient solution for automated LUS interpretation, with improved boundary precision.

In-context learning (ICL) enables generalization to new tasks with minimal labeled data. However, mainstream ICL approaches rely on a gridding strategy, which lacks the flexibility required for vision applications. We introduce Temporal, a time-contrastive self-supervised objective that pretrains a prompt retriever for visual ICL, and formulate ICL as a video object segmentation (VOS) task. Temporal addresses key limitations of grid-based methods that restrict the number and resolution of context images. By reframing ICL as a VOS problem, our approach supports a variable number of context images while preserving their full resolution. To address the challenge of selecting optimal context sets for queries, we pretrain a prompt retriever on videos via self-supervised learning, where adjacent frames serve as positives and distant frames as negatives. For image segmentation, the prompt retriever selects relevant sequences that, when combined with the query, form coherent videos for VOS processing. For video segmentation, it identifies keyframes, predicts their masks using our ICL pipeline, and propagates them throughout the sequence. When evaluated on MICCAI FLARE 2022, our method achieves substantial improvements over baselines: 90.95% Dice score for image segmentation (10.64% improvement) and 92.45% Dice for video segmentation (14.88% improvement).