The study presents a weakly supervised deep learning approach to automatically classify primary liver cancers (PLCs) from routine-stained biopsies. The researchers selected 166 PLC biopsies, divided into training, internal validation, and external validation sets. Whole-slide images were annotated by pathologists, and tiles were extracted for training a ResNet18 neural network. An unsupervised clustering algorithm was then applied to identify specific features of hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (iCCA). The model achieved 100% diagnostic agreement with pathological diagnosis for HCC in the internal validation set and 78% for iCCA. For combined hepatocellular-cholangiocarcinoma (cHCC-CCA), the model identified tiles within a slide, which could facilitate the diagnosis of cHCC-CCA. The study highlights the potential of weakly supervised deep learning in extracting meaningful features from routine histopathology slides, particularly for challenging cases like cHCC-CCA.The study presents a weakly supervised deep learning approach to automatically classify primary liver cancers (PLCs) from routine-stained biopsies. The researchers selected 166 PLC biopsies, divided into training, internal validation, and external validation sets. Whole-slide images were annotated by pathologists, and tiles were extracted for training a ResNet18 neural network. An unsupervised clustering algorithm was then applied to identify specific features of hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (iCCA). The model achieved 100% diagnostic agreement with pathological diagnosis for HCC in the internal validation set and 78% for iCCA. For combined hepatocellular-cholangiocarcinoma (cHCC-CCA), the model identified tiles within a slide, which could facilitate the diagnosis of cHCC-CCA. The study highlights the potential of weakly supervised deep learning in extracting meaningful features from routine histopathology slides, particularly for challenging cases like cHCC-CCA.