Medical imaging generates massive datasets, from gigapixel pathology slides to high-resolution 3D CT scans. While compression is a logistical necessity for storage and transmission, traditional lossy algorithms often prioritize visual smoothness over clinical accuracy. This creates a "diagnostic-fidelity tradeoff": as compression ratios increase, the resulting artifacts can either obscure subtle pathologies or, worse, mimic them.
The Artifact Problem
Standard lossy compression, such as JPEG, works by discarding high-frequency information that the human eye typically ignores. However, in medical contexts, these high-frequency details often contain the primary diagnostic signal. Fischer et al. observe that JPEG artifacts like blurring and blockiness can obscure cellular textures or nuclear morphology essential for pathology.
The danger is not just the loss of information, but the introduction of "technical noise" that misleads observers. Yip et al. demonstrate that high levels of compression introduce pixelation artifacts that deep learning models frequently misidentify as pathological features, leading to a surge in false positives. This suggests the existence of a diagnostic fidelity threshold: a quality level below which a model’s clinical reliability collapses because it begins "diagnosing" the compression noise rather than the anatomy.
Region-of-Interest Coding
One intuitive solution to the compression-fidelity tradeoff is to treat different parts of an image with varying levels of scrutiny. In a mammogram or a brain MRI, the vast majority of the pixels represent background air or non-critical tissue.
Luti et al. and Xu et al. propose Region-of-Interest (ROI) coding frameworks. By using Computer-Aided Diagnosis (CAD) or atlases to segment critical zones, these systems apply lossless compression to diagnostically relevant regions while allowing aggressive lossy compression for the background. This hybrid approach allows for high overall compression ratios—sometimes twice as efficient as standard JPEG2000—without altering a single bit of the data required for clinical interpretation.
Feature-Level vs. Pixel-Level Fidelity
Traditional compression metrics like Peak Signal-to-Noise Ratio (PSNR) measure fidelity by calculating the mathematical distance between original and reconstructed pixels. However, a reconstructed image can have a high PSNR while still losing the specific "deep features" that a diagnostic model needs.
Researchers are shifting toward feature-preserving compression. Kar et al. argue that for mammograms, preserving diagnostically relevant features is more important than achieving perfect pixel-wise reconstruction. Similarly, Fischer et al. introduced deep-learning-based compression that optimizes for "deep features," ensuring that cellular markers remain intact even when the raw pixel values change. This shift in perspective allows for much higher compression ratios because the algorithm only "protects" the information that contributes to the final diagnosis.
Bridging the Gap to Clinical Deployment
The transition from raw data to compressed formats often happens at the point of capture or during transmission over mobile networks. Doukas and Maglogiannis highlight that in telemedicine, network fluctuations necessitate adaptive coding to ensure critical data reaches clinicians.
However, the loss of metadata during these conversions can be just as damaging as the loss of pixels. Siegel and Nguyen et al. emphasize that converting high-fidelity DICOM files to lossy formats like JPEG strips away Hounsfield Units and bit-depth. This prevents clinicians and AI models from using "windowing" techniques to adjust contrast, effectively blinding them to subtle tissue differences.
To mitigate these effects in the real world, Antony demonstrates that training AI models on paired "original and compressed" datasets can reduce diagnostic errors by up to 80%. By teaching the model to recognize and ignore compression noise, researchers can maintain diagnostic integrity even when data must be transmitted through aggressive, lossy channels like mobile messaging apps.
Key Foundational Work
Adaptive Transmission of Medical Image and Video Using Scalable Coding and Context-Aware Wireless Medical Networks (2007) — contributed 1 concept elements to this direction.
Image resizing using saliency strength map and seam carving for white blood cell analysis (2010) — contributed 1 concept elements to this direction.
An interactive toolbox for atlas-based segmentation and coding of volumetric images (2007) — contributed 1 concept elements to this direction.