Identifying diseases using biopsy images remains a significant research challenge in Computer-Aided Diagnosis (CAD). In this study, we analyze and compare the performance of several modern transformer architectures for the automatic classification of liver cancer morbidity. Various layer unfreezing strategies are examined, along with the potential of using self-supervised Masked Autoencoders (MAE). The findings reveal that fine-tuning too many layers can result in overfitting, while restricting training to only the final classification layer and the layer normalization layers achieves the best results. Additionally, the study highlights the critical importance of pretraining data in determining the final performance of MAE models. As expected, pretraining and downstream tasks performed on the same dataset yield the highest accuracy. Furthermore, pretraining on images from a similar domain significantly improves performance compared to using a generally pretrained model, underscoring the benefits of domain-specific pretraining for CAD applications.