Diagnosis-oriented learned image compression system for chest radiographs with Regions-of-interest control

Abstract

This thesis presents DiagLIC, a diagnosis-oriented learned image compression (LIC) framework specifically designed for chest radiographs. Unlike conventional codecs and general-purpose LIC models that prioritize visual fidelity or bitrate efficiency alone, the proposed method in- corporates both diagnostic awareness and region-of-interest (ROI) preservation directly into the compression objective. Leveraging a transformer-based architecture with prompt-conditioned Swin Transformer blocks, the system supports variable-rate compression through a single model and allocates higher representational capacity to clinically significant regions such as lesions, opacities, or cardiomegaly zones. The framework is jointly optimized using a composite loss function that integrates rate-distortion and classification performance, ensuring that compressed images retain features essential for downstream diagnosis. Empirical evaluation on NIH and VinDr-CXR subsets demonstrates that DiagLIC significantly outperforms classical codecs (e.g., JPEG, WebP, JPEG2000) and state-of-the-art learned compression baselines in both PSNR and diagnostic accuracy. Across six quality levels, DiagLIC consistently achieves higher PSNR (e.g., 40.42 dB at 0.0951 bpp) and yields an average weighted PSNR of up to 52.49 dB, with ROI-specific PSNRs exceeding NROI PSNRs by over 1 dB at all levels. For diagnostic classifi- cation, DiagLIC maintains high fidelity, with only a minimal drop in average AUC (0.014) com- pared to uncompressed images, achieving AUC scores above 0.70 across all 8 tested patholo- gies and no degradation in certain categories such as Nodule/Mass. Statistical analysis confirms that the improvements in rate-distortion performance over baselines are significant, while the reduction in diagnostic performance remains clinically negligible. Qualitative comparisons fur- ther reveal that DiagLIC effectively suppresses artifacts and preserves lesion-level details, even at aggressive compression. This work demonstrates the feasibility and clinical value of inte- grating semantic objectives into image compression models, laying the groundwork for future diagnosis-preserving compression techniques in medical imaging.