The rise of the Internet of Medical Things (IoMT) has introduced an unprecedented volume of networked medical devices, which in turn has raised serious cybersecurity concerns. Intrusion Detection Systems (IDS) are crucial for safeguarding patient data and device functionality in these environments. However, most existing IoMT IDS solutions rely on conventional Machine learning (ML) or Deep learning (DL) models that demand extensive training data and computational resources. These heavy approaches are often trained to specific network protocols and evaluated on generic or outdated datasets, such as KDD, NSL-KDD, that do not reflect modern IoMT traffic. In this paper, we propose a resource-efficient Few-Shot Learning (FSL) framework for IoMT intrusion detection. Our model integrates a Cosine Prototypical Network with Adaptive Temperature Scaling (ATS) and a Prototype Spread Regularizer (PSR) to enable rapid adaptation to new attack patterns using minimal data. We leverage a new multi-protocol IoMT security dataset (CICIoMT2024) that contains realistic network traffic from healthcare devices under various cyberattacks. that our approach achieves 99.05% binary detection accuracy and in comparsion to other baseline models performance under similar conditions, our model requires 28× less training data and consumes 13× less peak memory (1.3MB) than standard baselines, while outperforming CNN and TabNet architectures. These results establish our proposed few-shot framework as a feasible, lightweight solution for adaptable security monitoring in critical healthcare networks.