Hello - You have encountered a major design choice in this challenge, and what we believe makes this challenge instructive, real-world, and difficult (although we are not blind to the leaderboard 😄 - it is incredible how well a community can do when collaborating, better than any one individual team).
\nA related discussion is here. It outlines the design principles that led to perceived inconsistencies.
\nBelow, I will focus on specific questions.
\nWere the labels assigned independently from the MRI images, rather than extracted from the reports?
\nYes
\nIf image interpretation and report text disagree, should the image-derived label be considered authoritative?
\nYes. Note that only a small sample of provided data contains both. It is intended to help participants surface this conclusion.
\nCould you provide the exact positive-label definitions for osteoarthritis versus chondropathy, bone contusion versus nonspecific bone marrow edema, acute fracture versus osteochondral/subchondral/insufficiency fractures, and joint effusion versus synovitis?
\nPlease refer to the pinned Overview discussion post. At the bottom, textual and image-based explanations of each abnormality is provided.
\nDo negative labels mean that the finding was confirmed absent, or can they also mean that it was not annotated?
\nIt means the finding was annotated as absent using methods outlined in Overview discussion post under Label Description section.
\nAre multiple reports or bilateral examinations sometimes associated with one StudyInstanceUID? If so, how were the study-level labels consolidated?
\nYes. In clinical practice, both knees may occasionally be scanned under one StudyInstanceUID. For the challenge, each bilateral study or bilateral report was individually reviewed, and the released report text or DICOM metadata was adjusted as needed to provide sufficient information for participants to disambiguate the relevant study/studies.
\nShould the examples above be considered expected report-image disagreement, or are any of them possible annotation issues?
\nDiscrepancies are plausible and expected because clinical reports typically involve one signing radiologist who created it for clinical care, and the image-based labels uses multiple readers with stricter image-based thresholds.
\nI hope this helps, and good luck on the challenge!
" }, { "id": 3510426, "authorName": "Handudu", "votes": 0, "postDate": "2026-08-08T13:40:09.350000", "content": "Thank you; this is crucial for the competition.
" } ], "index": { "id": "733826", "title": "Possible inconsistencies between MRI reports and provided labels", "authorName": "", "commentCount": "2", "votes": "27", "postDate": "2026-08-08 13:08:26.562000" }, "competition": "rsna-knee-abnormality-detection" }