Bone fenestration and dehiscence (FD) are conditions where the bone surrounding a tooth is lost or damaged, potentially leading to complications such as gum recession or infection. Accurate identification of FD is essential for proper dental care and treatment planning. Cone-beam computed tomography (CBCT) provides detailed 3D images to identify FD but is costly, exposes patients to radiation, and has limited availability. In contrast, intraoral images, commonly used due to their accessibility and affordability, present challenges in detecting FD, as the condition can appear subtle and easily overlooked. Existing supervised methods, such as FD-SOS, leverage multi-task Vision-Language Models (VLMs) that are fine-tuned on public dental datasets. By integrating conditional contrastive denoising (CCDN) with teeth-specific matching assignments, these methods enhance generalization while prevent overfitting. However, labeling data for FD detection is time-consuming and costly, making it challenging to obtain large-scale, high-quality FD datasets. To address these challenges, we propose FD-SSD, a semi-supervised framework that leverages a teacher-student architecture with an adaptive query strategy derived from pseudo-labels and multi-scale input features, along with a query filtering mechanism. By utilizing both labeled and unlabeled data, FD-SSD reduces reliance on pre-training and extensive annotations. Our approach outperforms the supervised pre-trained DINO, boosting mAP from 62.08% to 68.3%. Additionally, FD-SSD achieves a significant performance improvement over FD-SOS, increasing mAP from 65.67% to 68.3%. These results demonstrate its effectiveness in FD detection, even with limited labeled data. Code is available at: https://github.com/tahirashehzadi/FD-SSD