
%Aigaion2 BibTeX export from HES SO Valais Publications
%Thursday 19 November 2026 09:52:02 PM

@ARTICLE{,
    author = {Damm, Hendrik and Reinartz, Lea and Bracke, Benjamin and Eryilmaz, Bahadir and Nath, Praveen and Br{\"{u}}ngel, Raphael and Schmidt, Cynthia S. and Sch{\"{a}}fer, Henning and Ben Abacha, Asma and de Herrera, Alba G. Seco and M{\"{u}}ller, Henning and Friedrich, Christoph},
  keywords = {computer vision, Image Captioning, ImageCLEF, Multi-label classification, Radiology, Synthetic Data, Vision-Language Models},
     month = oct,
     title = {Overview of ImageCLEFmedical 2026 – Medical Concept Detection and Caption Generation with Synthetic Data Extensions, CLEF working notes, Jena, Germany, 2026.},
   journal = {Coeur},
    volume = {4283},
      year = {2026},
     pages = {3122-3145},
      note = {Actes de conf{\'{e}}rence: CLEF 2026 Working Notes},
  abstract = {The 10th Edition of the ImageCLEFmedical Caption task extends the established concept detection and caption
prediction benchmark with two synthetic-data subtasks and a manually evaluated explainability task. The
2026 release is built from finalized 2025 outputs and expanded with new PubMed Central data and updated
annotations. The main dataset contains 131853 images (97364 train, 19240 validation, 15249 test) and 2646
unique UMLS Concept Unique Identifiers (CUIs). The synthetic-caption dataset contains 131723 images (97222
train, 19 239 validation, 15262 test) and 2174 unique CUIs. Both datasets were harmonized through missing-image
pruning, zero-concept removal, and CUI-to-name mappings. The final test phase received 102 finished runs across
four prediction leaderboards, and the explainability task received three manually evaluated submissions. Best
hidden-test scores were 0.5790 F1-score for main concept detection, 0.3755 overall for main caption prediction,
0.5609 F1-score for synthetic concept detection, and 0.5840 overall for synthetic caption prediction. The best
explainability submission reached an overall manual score of 3.69 on a 1–5 scale. This paper documents the 2026
task design, data preparation, synthetic-data extension, evaluation setup, organizer baselines, and final results}
}

