JPEG Compression and False Positives in Medical AI

Lossy JPEG compression can indeed introduce false positive anomalies in computer-aided medical diagnostics (CAD). While lossy compression effectively reduces file sizes for storage and transmission, it achieves this by permanently discarding high-frequency image data. In sensitive medical imaging domains—such as digital mammography, histopathology, and chest radiography—the visual artifacts generated by this compression can mimic genuine pathological features, leading diagnostic algorithms to misinterpret algorithmic noise as clinical abnormalities.

How Lossy JPEG Creates Artifacts

The standard JPEG algorithm divides an image into 8x8 pixel blocks and applies a Discrete Cosine Transform (DCT) to convert spatial data into frequency components. During the quantization step, the algorithm discards high-frequency details that the human eye typically does not notice.

However, this process introduces two primary types of structural distortions:

• Blocking artifacts: Sharp, artificial boundaries along the edges of the 8x8 blocks.
• Ringing and blurring: Ripples or halos around sharp edges and loss of fine-textured details.

Impact on Computer-Aided Diagnostic (CAD) Algorithms

Unlike human clinicians, who can often subconsciously filter out compression artifacts based on contextual anatomy, machine learning models analyze pixel-level gradients and local textures.

1. Feature Mimicry: Blocking artifacts and high-frequency ringing can structurally resemble microcalcifications in mammograms, tiny lung nodules in chest X-rays, or microaneurysms in retinal fundus photography. A CAD system evaluated on compressed imagery can easily classify these sharp edge transitions as suspicious lesions.
2. Texture Alteration in Deep Learning: Convolutional neural networks (CNNs) rely heavily on subtle textural cues to differentiate benign tissue from malignant growths. Lossy compression smooths micro-textures, potentially causing models to misread tissue density variations or miscalculate margins, triggering a false positive classification.
3. Domain Shift: Most medical AI models are trained on pristine, uncompressed DICOM datasets. Deploying these models on compressed images introduces an out-of-distribution shift. Even subtle compression ratios can cause a significant drop in model specificity, inflating the false positive rate.

Clinical Implications and Industry Standards

In a diagnostic setting, false positives lead to patient anxiety, unnecessary follow-up imaging, invasive biopsies, and increased healthcare costs. Consequently, regulatory frameworks and medical imaging standards—such as DICOM guidelines issued by the American College of Radiology (ACR) and the Radiological Society of North America (RSNA)—strictly regulate or prohibit irreversible (lossy) compression in primary diagnostic workflows.

While lossy compression may be acceptable for secondary review, telemedicine triage, or patient portals, computer-aided diagnostic pipelines should utilize original raw data or mathematically lossless compression methods (such as JPEG-LS or Lossless JPEG 2000) to ensure diagnostic reliability and eliminate artifact-induced false positives.