Background: Magnetic Resonance Imaging (MRI) is the primary imaging modality for brain tumor diagnosis due to its rich soft-tissue contrast. However, the limited availability of annotated MRI datasets poses a major challenge for developing reliable deep learning models. In contrast, Computed Tomography (CT) scans are more abundant and easier to annotate, but their distinct imaging characteristics hinder direct knowledge transfer to MRI-based analysis. Existing cross-modality learning approaches often rely on paired data or annotated target-domain samples, which are impractical in clinical settings. Methodology: This work proposes a self-distilled cross-modality representation learning framework that enables effective knowledge transfer from CT to MRI without requiring paired scans or MRI annotations. The framework incorporates a bidirectional CT-to-MRI translation module to reduce low-level appearance discrepancies between modalities. A modality-invariant embedding space is learned through contrastive cross-modality alignment, promoting structural consistency across domains. To provide stable and reliable supervision, a self-distillation mechanism is introduced in which an exponential-moving-average CT encoder guides MRI feature refinement. Additionally, a structure-aware attention module emphasizes tumor-relevant anatomical regions, enhancing the representation of small or low-contrast lesions. Results: The proposed method was extensively evaluated on BraTS-2018, BraTS-2020, TCIA-GBM, and CQ500 datasets. Experimental results demonstrate that the framework achieves state-of-the-art performance using only CT supervision, outperforming existing cross-modal and self-supervised baselines. Significant improvements were observed in precision, recall, and Hausdorff distance, highlighting the effectiveness of the proposed self-distilled learning strategy. Conclusion: The proposed self-distilled cross-modality framework offers a scalable and annotation-efficient solution for brain tumor analysis. By effectively transferring structural knowledge from CT to MRI and reducing modality discrepancies without paired data, the method improves robustness and clinical reliability.
