The abstract is read at the publisher; the summary is JClub's.
ARIPA achieves nuclear segmentation on hematoxylin and eosin images of formalin-fixed, paraffin-embedded patient-derived cancer organoids, reaching image-macro F1 0.9849 after expert feedback at IoU ≥ 0.5.
- Why it matters: Nuclear segmentation in PDOs is hard because morphology, staining, scale, and acquisition format vary, and existing methods often need training or lack robust human-in-the-loop refinement.
- What they did: The HITL system uses a vision-language model to plan, verify, and refine pre-trained segmentation tools, with tissue-aware tiling and image-derived calibration plus optional expert corrections. It was evaluated on 19 organoid images (more than 50,000 nuclei) across breast, head and neck, and lung.
- The result: After feedback, image-macro F1 rose from 0.8380 to 0.9849 (mean paired F1 increase 0.1468 in 17/19 images), and mean area under the F1-vs-IoU curve improved to 0.9700 versus 0.4317 for HoVer-Net.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.