A new AI framework diagnoses liver cancer types from routine tissue slides with accuracy rivaling more expensive laboratory staining methods.
Researchers trained the nnU-Net architecture on a publicly available dataset of 170 whole-slide images covering three liver cancer types: hepatocellular carcinoma, cholangiocellular carcinoma, and colorectal metastatic adenocarcinoma. Rather than classifying image patches, the model segments every pixel and then assigns a diagnosis based on whichever cancer type dominates the output. Four pathologists independently annotated the training images. Five-fold cross-validation produced balanced accuracy scores of 0.975, 0.950, and 1.000 for the three cancer types respectively.
The result matters because H&E staining is the default first step in most pathology labs — cheap, fast, and universal. Immunohistochemical staining, which the model performance matches or exceeds, is more expensive and adds turnaround time. If a model can reliably triage cases at the H&E stage, labs could reserve the costlier tests for genuinely ambiguous cases rather than running them routinely.
The dataset is small — 89 patients total — and the model has not been tested outside the controlled cross-validation setting, so clinical deployment is a longer road than the accuracy numbers suggest.