A weakly supervised deep learning method was used to classify primary liver cancers (PLCs) from routine tumour biopsies. The study aimed to automatically classify hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma (iCCA), and combined hepatocellular-cholangiocarcinoma (cHCC-CCA) using a two-cluster model. The model identified specific features of HCC and iCCA, but no specific features of cHCC-CCA were recognized. The model correctly predicted the diagnosis in 97% of HCC cases and 83% of iCCA cases. For cHCC-CCA, the proportion of HCC and iCCA tiles varied widely. The model's predictions showed high agreement with pathological diagnoses, particularly for HCC (100% agreement in internal validation and 96% in external validation) and iCCA (78% and 87% agreement, respectively). The model's ability to identify HCC and iCCA tiles within a slide could aid in diagnosing cHCC-CCA. The study highlights the potential of weakly supervised learning in improving the diagnosis of PLCs, especially in challenging cases like cHCC-CCA. The method involved training a ResNet18 neural network on HES-stained biopsy images, using tumour/non-tumour annotations for supervision. An unsupervised clustering algorithm was then applied to classify the tiles into two clusters, representing HCC and iCCA. The model's performance was validated on internal and external datasets, showing promising results in accurately classifying PLCs. The study underscores the importance of developing automated AI methods for biopsy specimens, as they can overcome limitations in traditional histological analysis and improve diagnostic accuracy.A weakly supervised deep learning method was used to classify primary liver cancers (PLCs) from routine tumour biopsies. The study aimed to automatically classify hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma (iCCA), and combined hepatocellular-cholangiocarcinoma (cHCC-CCA) using a two-cluster model. The model identified specific features of HCC and iCCA, but no specific features of cHCC-CCA were recognized. The model correctly predicted the diagnosis in 97% of HCC cases and 83% of iCCA cases. For cHCC-CCA, the proportion of HCC and iCCA tiles varied widely. The model's predictions showed high agreement with pathological diagnoses, particularly for HCC (100% agreement in internal validation and 96% in external validation) and iCCA (78% and 87% agreement, respectively). The model's ability to identify HCC and iCCA tiles within a slide could aid in diagnosing cHCC-CCA. The study highlights the potential of weakly supervised learning in improving the diagnosis of PLCs, especially in challenging cases like cHCC-CCA. The method involved training a ResNet18 neural network on HES-stained biopsy images, using tumour/non-tumour annotations for supervision. An unsupervised clustering algorithm was then applied to classify the tiles into two clusters, representing HCC and iCCA. The model's performance was validated on internal and external datasets, showing promising results in accurately classifying PLCs. The study underscores the importance of developing automated AI methods for biopsy specimens, as they can overcome limitations in traditional histological analysis and improve diagnostic accuracy.