{
  "id": 750103,
  "title": "Threshold selection and honest validation in multi-label imaging",
  "url": "/competitions/grand-xray-slam-division-b/discussion/750103",
  "author_name": "Elmaddin Guliyev",
  "post_date": "2026-10-16T22:34:18.095000",
  "votes": 0,
  "comment_count": 0,
  "views": null,
  "content": "<p>AUROC and F1 answer different questions: AUROC measures ranking, while F1 also depends on the decision threshold. If class thresholds are tuned on the same cases used to report F1, the reported score can be optimistic. A stronger evaluation keeps threshold selection inside the validation process and reports results on untouched patient groups. Per-class results alongside macro averages also help show how rare classes behave. This is a general validation note, not a result from the active RSNA competition.</p>",
  "messages": [
    {
      "id": 3540011,
      "postDate": "2026-10-16T22:34:18.097Z",
      "content": "<p>AUROC and F1 answer different questions: AUROC measures ranking, while F1 also depends on the decision threshold. If class thresholds are tuned on the same cases used to report F1, the reported score can be optimistic. A stronger evaluation keeps threshold selection inside the validation process and reports results on untouched patient groups. Per-class results alongside macro averages also help show how rare classes behave. This is a general validation note, not a result from the active RSNA competition.</p>",
      "rawMarkdown": "AUROC and F1 answer different questions: AUROC measures ranking, while F1 also depends on the decision threshold. If class thresholds are tuned on the same cases used to report F1, the reported score can be optimistic. A stronger evaluation keeps threshold selection inside the validation process and reports results on untouched patient groups. Per-class results alongside macro averages also help show how rare classes behave. This is a general validation note, not a result from the active RSNA competition."
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3540011": "AUROC and F1 answer different questions: AUROC measures ranking, while F1 also depends on the decision threshold. If class thresholds are tuned on the same cases used to report F1, the reported score can be optimistic. A stronger evaluation keeps threshold selection inside the validation process and reports results on untouched patient groups. Per-class results alongside macro averages also help show how rare classes behave. This is a general validation note, not a result from the active RSNA competition."
  }
}