Medical image segmentation is crucial in modern medical diagnostics, especially in accurately locating and identifying brain tumors. However, current segmentation models are limited by the availability of high-quality labeled data due to privacy and ethical constraints, as well as the scarcity of brain tumor datasets. To address this issue, we propose an innovative pre-training framework that integrates an Intelligent Plug-and-Play Module (IPM) into existing two-stage segmentation tasks, thereby significantly enhancing model performance in data-scarce scenarios. The IPM includes a semantic extractor and a pixel optimizer, utilizing self-distillation and self-modeling techniques to process masked patches, accelerating convergence and promoting efficient representation learning. Through this integrated approach, we successfully improved model performance in downstream tasks. Downstream experiments on well-known 3D brain tumor segmentation benchmarks validated the effectiveness of our method. Models equipped with the IPM consistently achieved high Dice coefficients and HD95 scores on the BraTS 2021 and MSD 2019 Task-01 Brain Tumor datasets. This study underscores the transformative potential of integrating the IPM into medical image segmentation in scenarios with sparse labeled data.