约翰霍普金斯与斯坦福发布 Merlin-Cancer-Super,9 器官肿瘤分割超 VLM
MaziyarPanahi · x · 2026-11-04
Stanford and Johns Hopkins teams released Merlin-Cancer-Super on Hugging Face, built on R-Super (MICCAI 2025 Best Paper Award, top 2). It surpasses state-of-the-art public AI models and leading VLMs at detecting and segmenting tumors across 9 organs: bladder, gallbladder, spleen, esophagus, stomach, duodenum, prostate, uterus, adrenal glands.
Method highlights:
- R-Super turns radiology reports into per-voxel supervision: an LLM extracts tumor info from reports first
- New Volume Loss & Ball Loss losses push segmentations to match reports in tumor count, diameters, and locations
Trained on Merlin Plus (Stanford, 25,494 CT scans with longitudinal metadata and per-voxel annotations), openly available on HF. Research on public data, not medical advice.
「多模态」频道最新
- Adobe 与 KAIST 发布 MoviEdit：一句话或单帧即可一致重渲多镜头长视频 — _akhaliq · 2026-11-04
- Fable 5.5 旧 checkpoint 生成的 voxel 宝塔花园,被赞迄今最佳效果 — legit_api · 2026-11-04
- UIST 2026 论文 CinemaWorld:用 LLM 与 3D 场景生成做电影增强现实 — ryosuzk · 2026-11-04
- HYPER3D Rigging 发布:3D 绑定可用提示词编辑,带250+动作预设 — _akhaliq · 2026-11-04
- 开源 ComfyUI-StudioLight v1.0：像摄影师一样控制生成图布光 — Neat_Two_8536 · 2026-11-04
- 用 Gemini 写代码绘制《摩诃婆罗多》第一部分，每帧均由代码渲染 — dx8xb · 2026-11-04