Accurate brain tumor segmentation in MRI is critical, yet deep learning requires costly annotations. Semi-supervised learning (SSL) reduces this burden but still depends on initial labeled data to bootstrap training. We propose a hybrid framework combining a training-free analytical stage with an SSL network to reduce this dependency.

Our two-stage framework includes: (1) APP-RSC, a training-free algorithm generating coarse masks by iteratively pruning non-tumor voxels via intensity statistics. With a relaxed convergence threshold and morphological dilation, it achieves 98.3% sensitivity without ground-truth annotations. (2) A Mean Teacher semi-supervised 3D U-Net that refines these masks using only 6% labeled data and APP-RSC-generated pseudo-labels, optimized via supervised, pseudo-supervised, and consistency losses.

Evaluated on BraTS2023, APP-RSC processes a volume in ∼1.2 s on CPU. The Mean Teacher refinement significantly improves the Dice score from 0.17 to 0.825 (median 0.852) on the test set, with 99.3% specificity (Wilcoxon p=1.19×10 ^-7 ). This performs competitively against state-of-the-art SSL approaches using larger labeled datasets.

This framework bridges interpretable analytical segmentation and high-performance deep learning. Crucially, it drastically reduces the bootstrap dependency of existing SSL methods by requiring zero annotations for initial mask generation. This offers a practical, transparent, and annotation-efficient solution, particularly advantageous for resource-constrained clinical environments.