Answers are grounded in this paper and cite their sources.
Burned-in protected health information (PHI) in DICOM pixel data blocks multi-site imaging research, and better recognition is assumed to be the route to removal. We asked whether detection or the PHI decision limits removal. A layered pixel pipeline built on Florence-2 OCR was evaluated on MIDI-B with the CBIIT validation script. Against the answer key’s OCR-independent coordinates, two controlled arms swapped the engine with the decision fixed, then varied the decision over detections. A nine-condition degradation sweep ran to redaction, with Wilson intervals and exact McNemar tests. On clean images, a 0.23 B vision-language model and a CNN OCR at library defaults detected 34/35 and 35/35 regions, flagged the same 10/35 with the decision fixed, and removed the same 33/35. Over fixed detections, Microsoft Presidio flagged 12/35 and a generic rule set 10/35, as 25 of 35 regions are unpatterned markers; redacting every detection reached 35/35, so both shipped misses were decision misses. Under degradation, both engines detected the text in every condition; removal differed most at blur radius 4, 32 versus 26 ( p = 0.07), where the CNN OCR misread short markers (exact match 17/25 versus 10/25). An earlier-reported collapse to 18/35 came from a confidence floor in our harness. The strict profile scored 33/35 at 86.7% series accuracy, and the retain profile 32/35 at 97.6%. Removal on clean text is limited by the PHI decision, not the detector. Under synthetic degradation, the engines differ in transcribing short strings, not in detecting them, by a margin this sample cannot resolve. Results are benchmark- and engine-specific.
No takes yet. Share an insight, caveat, or question.
D. Patil (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: