1*Department of Computer Science and Engineering, Bharath Institute of Science and Technology (BIST), 173, Agaram Road, Selaiyur, Tambaram, Chennai-600073, Tamil Nadu. Email: sindhudevicse24@gmail.com
2Department of Computer Science and Engineering, Bharath Institute of Science and Technology (BIST), 173, Agaram Road, Selaiyur, Tambaram, Chennai-600073, Tamil Nadu. Email: godlinatlas.cse@bharathuniv.ac.in
Medical image analysis systems often suffer from significant performance degradation when deployed across hospitals, scanners, or acquisition protocols that differ from the training data. This domain shift [5] poses a major barrier to real-world clinical adoption of deep learning [1][8] models. In this paper, we propose a novel Domain-Invariant Medical Representation Learning [18][19] Framework [DIMRL] that explicitly disentangles task-relevant anatomical features from domain-specific acquisition characteristics. The proposed framework employs adversarial domain alignment, regularity-centric feature regularisation, and uncertainty-aware fusion to enable reliable generalisation across unseen clinical domains in the absence of target-domain labels. Experimental results on multi-centre medical imaging datasets [16] show significant gains in cross-domain performance, reduced dependence on scanner-specific characteristics, and preservation of clinically informative features over conventional domain adaptation [2][4] baselines. This study demonstrates that domain-invariant feature learning is essential for scalable and trustworthy medical image analysis [15][16].
Keywords: Domain adaptation, Medical Image Analysis, Domain Generalisation, Robust Learning, Domain-Invariance.
How to cite this article: Devi AS, Atlas LG. Domain-Invariant Representation Learning for Robust Medical Image Analysis Across Clinical Sites. Int J Drug Deliv Technol. 2026;16(15s): 369-374. DOI: 10.25258/ijddt.16.15s.43
Source of support: Nil.
Conflict of interest: None