Heavy JPEG Compression Made a Dental AI System Miss Radiographic Findings

Edited by Dentaltown staff

Saving panoramic radiographs at heavy JPEG compression caused a commercial dental artificial intelligence system to miss findings it had flagged on uncompressed versions of the same images, according to a study published online Sept. 12 in the Journal of Dentistry.

Investigators saved 100 panoramic radiographs as uncompressed TIFF files, which served as the reference standard, and at five JPEG compression levels: 90, 70, 50, 30, and 10. All 600 images were processed by a commercial system, X-ray Insights from dentalXrai, which assigned binary labels across nine diagnostic categories at 32 tooth positions per image. Agreement between each compressed image and the system’s own uncompressed output was measured with Cohen’s kappa, with 0.90 set in advance as the threshold.

Moderate compression held up well. Every label showed near-perfect agreement at quality levels 90 and 70, at kappa of 0.920 or higher. Four prosthetic and restorative labels — implant, bridge, crown, and root canal filling — were unaffected at every compression level, though implants and bridges appeared at 1.3% of positions or fewer.

Five labels fell below the threshold as compression increased. Periapical radiolucency and filling dropped at quality 30, mandibular canal at quality 50, and caries and present or intact teeth at quality 10. At the heaviest compression, positive percent agreement for periapical radiolucency fell to 0.614 and per-tooth exact agreement to 89.1%.

The failures were omissions rather than inventions. Compression mainly caused findings to disappear, though the system occasionally registered tooth positions that were not there.

Aleš Fidler and Veronika Krenker of the University of Ljubljana and the University Medical Centre Ljubljana conducted the study with Mohmed Isaqali Karobari of Saveetha Dental College and Hospitals in Chennai and Akhilanand Chaurasia of King George’s Medical University in Lucknow. The Government of the Republic of Slovenia funded the work, and the authors declared no competing interests.

The authors recommended a minimum compression level of 50 and said the compression applied to an image should be reported alongside it. They noted that roughly 90% of the available file-size reduction is achieved at quality 90 with negligible loss of agreement.

They also cautioned that the results describe the output stability of one system rather than its diagnostic accuracy, and should not be generalized to other AI products without validation.

Sources:
Journal of Dentistry, “Effect of Lossy JPEG Compression on AI Diagnostic Output Stability in Panoramic Radiograph Analysis,” published online Sept. 12, 2026: doi.org/10.1016/j.jdent.2026.107033
PubMed record, PMID 42731775: pubmed.ncbi.nlm.nih.gov/42731775