Back to all papers

Effect of Lossy JPEG Compression on AI Diagnostic Output Stability in Panoramic Radiograph Analysis.

September 12, 2026pubmed logopapers

Authors

Fidler A,Krenker V,Karobari MI,Chaurasia A

Affiliations (4)

  • Department of Endodontics and Operative Dentistry, Faculty of Medicine, University of Ljubljana, Ljubljana, Slovenia; Department of Restorative Dentistry and Endodontics, University Medical Centre Ljubljana, Ljubljana, Slovenia. Electronic address: [email protected].
  • Department of Restorative Dentistry and Endodontics, University Medical Centre Ljubljana, Ljubljana, Slovenia; Community Health Centre Slovenj Gradec, Slovenj Gradec, Slovenia.
  • Department of Conservative Dentistry and Endodontics, Saveetha Dental College and Hospitals, Saveetha Institute of Medical and Technical Sciences, Saveetha University, Chennai, Tamil Nadu, India.
  • Department of Oral Medicine and Radiology, King George's Medical University, Lucknow, Uttar Pradesh, India. Electronic address: [email protected].

Abstract

To evaluate the effect of lossy JPEG compression on the output stability of a commercial AI system applied to panoramic radiographs. One hundred panoramic radiographs were saved as uncompressed TIFF (reference standard) and at five JPEG compression levels (CL) - 90, 70, 50, 30, and 10 (J90-J10). All 600 images were processed by a commercial dental AI system (X-ray Insights, dentalXrai GmbH) which assigned binary labels for nine diagnostic categories across 32 tooth positions per image. Agreement between each compressed image and the system's own uncompressed output was quantified by Cohen's kappa (κ), with κ = 0.90 predefined as the agreement threshold. Compression predominantly caused loss of findings rather than false positives, with rare hallucination of absent tooth positions. All labels showed near-perfect agreement at J90 and J70 (κ ≥ 0.920). Four prosthetic and restorative labels (implant, bridge, crown, root canal filling) were unaffected at all CLs, although implant and bridge prevalence was low (≤1.3%). Five labels fell below the agreement threshold: periapical radiolucency and filling at J30, mandibular canal at J50, and caries and present/intact at J10. Positive percent agreement for periapical radiolucency fell to 0.614 and per-tooth exact agreement to 89.1% at J10. CL ≥ 50 preserved AI output across all clinically important labels, whereas CL ≤ 30 caused meaningful degradation, particularly for periapical radiolucency and caries. A minimum CL of 50 is recommended, and the compression applied should be reported. These findings reflect output stability rather than absolute diagnostic accuracy and should not be generalized to other AI systems without validation. Most file-size reduction is achieved at CL 90 (≈90% reduction) with negligible loss of output agreement. Diagnostic tasks are not equally affected - low-contrast findings degrade whereas high-contrast ones remain robust. Degradation occurs predominantly through missed findings rather than false positives. A CL below 50 should therefore be avoided.

Topics

Journal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAISlice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.