Whole-slide pathology assistant that prunes uninformative tissue inside its Mamba–Transformer slide encoder before an LLM answers questions.
No providers recorded yet. Browse all providers
A whole-slide image tiled at diagnostic resolution yields thousands of patches, and most of them are normal stroma, fat or background that says nothing about the question being asked. Slide-level multimodal language models nonetheless tend to encode every patch densely and hand the full sequence, or a pooled summary of it, to the language model, where self-attention cost grows with the square of the visual token count. Where earlier systems compress at all, they do it after the slide encoder has finished, so the encoder never learns which regions matter.
SLICEChat, from Koç University with Hacettepe University, moves that decision inside the slide encoder. The encoder alternates Vision Mamba blocks, which propagate context across very long patch sequences in linear time, with Transformer blocks that add global attention once the sequence is shorter. Between stages, small learned routers score the contextualized tokens and remove the lowest-scoring regions of tissue, so each stage works on a smaller slide than the last and the language model receives a compact, diagnostically focused representation.
It belongs to the same family as SlideChat and WSI-LLaVA: a slide encoder bridged to a 7B-class LLM and instruction-tuned for slide-level visual question answering. It is an unrelated model from a different group than SlideChat despite the similar name. The work was released as an arXiv preprint in September 2026.
Slides are tiled into 512 × 512 patches at 20× with the TRIDENT pipeline and embedded by a frozen CONCH v1.5 patch encoder. The slide encoder has six stages of three Vision Mamba blocks plus one Transformer block, with 2D mixed RoPE in the attention layers. Training runs in three stages: masked-autoencoder pretraining of the encoder on 11,410 unlabeled TCGA slides with pruning disabled; contrastive training of the pruners, attention pooler and a text projection with the backbone frozen; and single-stage joint tuning of the projector and a Qwen2.5-VL-7B-Instruct language model on slide captions and VQA pairs. Two checkpoints were trained, one on SlideInstruct (4,181 slides, 175,734 VQA pairs) and one on the WSI-Bench training set (9,836 slides, 167,425 VQA pairs), each about 7.8 billion parameters.
On SlideBench VQA the SlideInstruct model reaches 79.84% overall accuracy on the TCGA cohort and 59.09% on the out-of-distribution BCNB breast biopsy cohort, against 74.95% and 54.04% for SlideChat and 76.51% and 53.24% for HistoSelect under the authors' re-evaluation. On open-ended WSI-Bench questions the WSI-Bench model scores 0.607 WSI-Precision and 0.776 WSI-Relevance overall, above WSI-LLaVA's 0.538 and 0.760, with an LLM judge (Gemini) in place of the original GPT-4o. Ablations show that keeping 20% of tokens beats keeping all of them on both cohorts. On a typical 11,948-patch slide the model uses 15.05 GiB peak memory and 2.0 s latency on an A100, close to WSI-LLaVA and below SlideChat.
SLICEChat answers multiple-choice and open-ended questions about an entire slide — tumor type, grade, morphological findings, clinical considerations — and generates slide-level descriptions, making it a research tool for computational pathology groups building or benchmarking slide-level assistants. Its pruning design is most relevant where inference cost matters, such as screening large slide archives. The released code covers encoder training, LLM training and VQA inference from TRIDENT-extracted features.
SLICEChat's contribution is architectural: it shows that compressing a slide during encoding, under language supervision, can improve accuracy rather than trading it away, and that state-space backbones outperform the LongNet encoders used by earlier slide-level assistants in this setting. All evaluation is on retrospective public benchmarks drawn largely from TCGA, and scores are reported from the authors' own re-evaluation of baselines; there is no prospective or clinical validation. The code repositories and the Hugging Face weights carry no license file.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.