Loading…
Histopathology vision-language model handling image patches and gigapixel slides in one 15B checkpoint, across classification, VQA, and captioning.
No providers recorded yet. Browse all providers
Computational pathology has largely been built in two disconnected halves. Patch-level models such as CONCH, UNI and Virchow2 learn rich representations of small tissue crops, while slide-level systems like Prov-GigaPath aggregate thousands of those crops into a whole-slide prediction. Language-capable pathology assistants inherited the same split: most operate only on patches because of input-size limits, and the handful that read whole slides do one narrow job such as report generation. The result is a proliferation of single-purpose checkpoints that cannot share what each has learned.
CPath-Omni, from Westlake University with collaborators at Zhejiang University, Harvard University, The Ohio State University and the University of the Chinese Academy of Sciences, collapses that split into a single 15-billion-parameter large multimodal model. One checkpoint performs classification, visual question answering, captioning and visual referring prompting at both patch and whole-slide scale. The work was published at CVPR 2025, and the same group's later CPathAgent reuses its vision encoder.
The language backbone is Qwen2.5-14B. Training runs in four stages: aligning CPath-CLIP to the LLM through a two-layer MLP on the 700,145-pair caption corpus, full fine-tuning on 351,871 patch instructions across 21 datasets, continued pretraining on 5,850 cleaned whole-slide reports with only SlideParser unfrozen, and a final mixed stage over 33,830 slide-level instructions blended with 15% patch data.
Across seven tasks and 42 datasets the model leads on 39. On the PathMMU visual question answering benchmark it reaches 72.4% on the tiny split and 72.2% overall, ahead of PathGen-LLaVA by 13.8 points and slightly above the 71.8% recorded for expert pathologists. On whole-slide VQA it scores 67.3% open-ended and 70.8% closed-ended against 20.5% and 35.5% for GPT-4o. Report generation reaches BLEU-1 of 33.7 versus 31.8 for HistGen. CPath-CLIP alone averages 83.2% zero-shot accuracy over nine classification datasets, compared with 79.7% for the strongest prior pathology CLIP.
The model is aimed at diagnostic support where a pathologist wants one assistant across the whole workflow: subtyping across eight TCGA cancer types, grading, captioning of microscope-captured fields, drafting slide reports, and answering follow-up questions about a marked region. Because patch and slide handling live in the same checkpoint, a site deploying it does not need to maintain a separate model per task or per scale.
CPath-Omni is a concrete demonstration that unified training beats task-specific fine-tuning in pathology even when slide-level data is two orders of magnitude scarcer than patch data, and CPath-CLIP has already been adopted as a vision encoder in follow-on work. Practical access is uneven, however. The CPath-CLIP weights are public only as delta weights against gated Virchow2, so reconstruction requires a separate license grant from Paige; the 15B language model itself has not been released; and the repository carries no license file, leaving reuse terms undefined. Reported benchmark figures also come from the authors' own 42-dataset assembly rather than an independent evaluation.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.