This diploma thesis investigates deep learning methods for histopathological image analysis, focusing on prostate cancer grading from digitized whole-slide images. The main part of the work uses the PANDA dataset for slide-level ISUP grade prediction through multiple-instance learning and features extracted by frozen pathology foundation models. Four encoders are compared, and models combining their aligned feature representations are proposed, including residual gated fusion and counterfactual routing regularization.

The best single-encoder model achieved a mean quadratically weighted kappa of 0.922 on PANDA. In external evaluation on SICAPv2, the best result was achieved by the model with counterfactual routing regularization, reaching QWK 0.884. Attention analysis further showed that the model generally assigned higher importance to tumour regions; however, neither attention weights nor routing weights should be treated as complete explanations of its decisions.

The results show that pathology foundation models combined with multiple-instance learning provide a strong basis for prostate biopsy grading. They also highlight the importance of external validation and verification of model behaviour outside the development dataset.