Model-Augmented Masked Autoencoder with Domain-Adaptive Denoising for Robust Brain Tumor MRI Analysis

This work presents a Model-Augmented Masked Autoencoder (MAE) framework, augmented with domain-adaptive denoising and multi-level feature fusion, for self-supervised representation learning in MRI-based brain tumor analysis. The proposed approach enhances feature robustness under limited annotation by learning structure-preserving representations through masked reconstruction while explicitly addressing scanner-induced variability. Unlike conventional MAE designs, the framework integrates a Domain-Adaptive Denoising (DAD) module and latent–reconstruction feature fusion, enabling improved robustness across heterogeneous MRI datasets and reducing feature variance arising from scanner and acquisition differences. Comprehensive ablation studies across multiple datasets demonstrate that MAE pretraining yields consistent performance improvements of approximately 1.0–1.4% in classification accuracy, with overall improvements ranging from approximately 2.7% to 3.7% across datasets when integrated into the complete framework. Qualitative and quantitative analysis further confirm effective suppression of structural noise while preserving diagnostically relevant features. Overall, the proposed framework improves representation stability, enhances cross-domain robustness, and provides a reliable foundation for self-supervised medical image analysis, supporting the development of generalizable and clinically applicable diagnostic AI systems.

@article{K2026,
  title     = {Model-Augmented Masked Autoencoder with Domain-Adaptive Denoising for Robust Brain Tumor MRI Analysis},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {7},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Indrakumar K and Ravikumar M and Mohammed A.S Al-mohamadi and Khalid N. R. Alharbi and Asma A. Alhashmi and Shouki A. Ebad and Abdulbasit A. Darem},
  doi       = {10.14569/IJACSA.2026.0170739},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170739}
}