Pathology AI has mostly worked on small patches of tissue. A research group at the Shanghai Artificial Intelligence Laboratory has now released a system designed to read entire slides.
The model, SlideChat, combines a patch-level encoder for cellular detail with a slide-level encoder that compresses a full specimen into one representation, then connects both to a pretrained language model. It answers clinical questions and drafts diagnostic-style reports across 31 cancer types, according to a study in Nature Cancer.
Training relied on SlideInstruction, a 274,233-sample multimodal instruction dataset pairing whole-slide images with pathology reports and Q&A pairs. Sources include The Cancer Genome Atlas, the Clinical Proteomic Tumor Analysis Consortium, the BCNB breast cohort and the HISTAI dataset, plus a controlled-access hepatobiliary cohort from the Eastern Hepatobiliary Surgery Hospital.
Testing ran on SlideBench, which covers five cohorts, 8,836 closed-ended questions, 129 open-ended ones and 3,149 whole-slide reports. SlideChat beat leading baselines by 19.1 percentage points on closed-ended accuracy and 7.7 points on report generation. Pathologists scored its open-ended answers highest on diagnostic accuracy.
The authors point to a bladder cancer case where patch-level models misread invasive stage while SlideChat reported the correct pT3 stage, an error class tied to reasoning that spans the whole slide.
Shanghai General Hospital and Stanford University School of Medicine contributed to the work. No clinical deployment timeline has been announced.
