Although deep learning–based models have achieved impressive performance in segmenting complex anatomical structures and pathological regions, their adoption in clinical practice is often limited by high computational demands and substantial memory requirements. To address these constraints, lightweight architectures such as UNeXt have recently gained attention, offering a favorable trade-off between efficiency and segmentation accuracy by replacing heavy convolutional backbones with tokenized MLP-based components. Nevertheless, despite their efficiency, lightweight models may struggle to capture fine-grained and class-specific features, particularly in scenarios involving limited annotated data and significant intra-class variability. In this work, we introduce an enhanced segmentation framework that incorporates prototype-based learning into the UNeXt architecture. By learning representative feature prototypes for each class, the proposed method exploits similarity relationships in feature space to guide pixel-wise predictions, thereby improving class discrimination and robustness in noisy or ambiguous regions. This prototype-guided refinement seamlessly augments the lightweight UNeXt pipeline without imposing substantial additional computational overhead, ensuring suitability for deployment in point-of-care and resource-constrained clinical environments. Experimental results demonstrate that the proposed approach consistently outperforms the baseline UNeXt model in terms of segmentation accuracy and boundary delineation, while preserving its lightweight characteristics. These findings underscore the effectiveness of prototype-enhanced lightweight architectures for practical medical imaging applications, including cancer screening, dermatological analysis, and ultrasound-based diagnostics. The code and datasets supporting this work are available from the authors upon request via email.