Mental health detection on social media is challenging due to limited labeled data, data imbalance, and informal text structures. This study proposes IS iPET, an incremental selection training strategy that enhances Pattern-Exploiting Training (PET) and iterative PET (iPET) by gradually incorporating and strategically selecting training samples for fine-tuning Masked Language Models (MLM). Additionally, a margin-based loss function improves class separability. Experiments on Chinese social media posts show IS iPET improves precision by 20% and F1-score by 10%, while maintaining strong performance with 50% less training data. In open-environment testing, IS iPET achieves 0.81 precision in help-seeking behavior detection, demonstrating its real-world applicability. These findings suggest IS iPET is an effective semi-supervised approach for mental health detection.