<p>Hyperspectral imaging has demonstrated great potential for medical image segmentation. However, the acquisition of precise pixel-level annotations is costly, which limits the availability of labeled data for supervised training. To address this issue, this paper proposes a prototype- and uncertainty-guided semi-supervised spectral-spatial Mamba framework, termed ProtoMamba-SS. The proposed framework consists of a Robust Teacher, a tri-map pseudo-label generation module, and a spectral-spatial Mamba Student. In the teacher branch, spectral attention, multi-scale feature extraction, and uncertainty estimation are introduced to derive stable semantic priors and reliable confidence information from labeled samples. Teacher predictions, class-wise prototype consistency, and uncertainty-aware filtering are jointly incorporated into the pseudo-label generation process, enabling the model to explicitly exclude ambiguous regions while producing confidence-weighted tri-map pseudo-labels. The student branch integrates a learnable spectral compression stem, a Spectral Mamba branch, a multi-scale Spatial Mamba branch with multi-directional scanning, spectral-spatial gated fusion, and dual segmentation and boundary heads, thereby jointly modeling spectral dependencies, spatial context, and contour details. Furthermore, an EMA-guided self-training strategy is adopted to stabilize pseudo-label propagation and iterative optimization. Experimental results on public datasets demonstrate that the proposed method achieves superior segmentation performance and more reliable boundary delineation compared with existing methods under limited annotation settings.</p>