Within the last decade, applications of deep learning in the computational pathology field have improved keeping up with the breakthroughs in artificial intelligence in recent years. When applied to cancer classification tasks, the proposed methods can aid the pathologists to detect the correct type of cancer more effectively and accurately. In the context of brain cancer, whole slide images (WSI) are usually used by expert pathologists to inspect the tissue of a tumor sample and determine the grade and type of tumors. The aim of this thesis is to implement and evaluate the state-of-the-art models for slide-level analysis on brain tumor classification. 

The used dataset (TCGA) includes 1484 WSIs from 748 subjects. The slides contain tissues extracted from different gliomas. A glioma is a type of malignant tumor which originates from the glial cells in the brain. The model must differentiate between the different types and grades of glioma, with a total of 5 possible classes. The implemented method includes a segmentation and patching module to segment the foreground and extract patches from each slide. Before training is performed, a feature extraction model produces features of 1024 elements for each of the patches. Two feature extraction models were used: a ResNet50 model pre-trained on the ImageNet dataset, and the foundation model UNI pre-trained on Mass-100K. CLAM is the model used for training, which is a combination of attention-based multiple instance learning (MIL) with clustering to enhance the training with positive and negative evidence. The model trained with ResNet50 features is referred to as RN-CLAM, and the one trained with UNI features as UNI-CLAM.

The performance of the model was assessed using the area under the curve (AUC) from the receiver operating characteristic (ROC) curve, precision, recall, F1-scores, confusion matrices and Matthews correlation coefficient (MCC). Uncertainty of the models was estimated using the Monte Carlo dropout method. A permutation test with 60 repetitions of the dataset split was performed to compare the models statistically (RN-CLAM vs. UNI-CLAM). Finally, a method is proposed to compute the correlation between CLAM attention and the cell density in the tissue, by using the attention maps generated with the trained CLAM models and the cell concentration maps generated using the QuPath program. The metrics used to compare the images are the Pearson correlation coefficient (PCC), structural similarity index measure (SSIM) and mean squared error (MSE).

UNI-CLAM achieved the best performance with all the introduced metrics with an average test AUC of 0.92 and an MCC of 0.61. RN-CLAM achieved a high performance as well, with an average AUC of 0.90 and an MCC of 0.54. Both models demonstrated moderate uncertainty, however, UNI-CLAM was more certain for most of the classes compared to RN-CLAM. Additionally, the models showed a higher uncertainty for the most relevant patches for the classification. The permutation test obtained a p-value of 0.0001 demonstrating that UNI-CLAM performed significantly better than RN-CLAM. Finally, the CLAM attention maps showed a relatively high correlation overall with the cell density maps, with UNI-CLAM (PCC=0.65, SSIM=0.69, MSE=0.07) showing a higher similarity than RN-CLAM (PCC=0.61, SSIM=0.69, MSE=0.08).

In conclusion, UNI-CLAM performed significantly better, was more certain about the predictions and showed a higher correlation with the cell density. The observed difference between both models was large enough to be proven statistically. The CLAM attention maps showed a high correlation with the cell density maps, but additional features within the tissue that influence the decision should be further investigated.