Abstract:
Deep learning has become a pivotal tool in computational pathology, particularly for classifying CNS tumors in histopathology images stained with hematoxylin and eosin (H&E). This study evaluates the performance of the Random Walker algorithm for segmentation and the Vision Transformer (ViT 32 B ) [6] model for classification tasks. The Random Walker algorithm demonstrated efficacy in handling unlabeled datasets with low computational requirements but faced challenges in scalability and manual intervention. Meanwhile, the ViT 32 B model ex- celled in capturing long-range dependencies and delivering robust classification with limited training samples, though it required substantial computational resources and large, high-quality datasets. [8] [4] To address these limitations, the study underscores the potential of hybrid approaches that combine the segmentation capabilities of Random Walker with the classification strength of Vision Transformers. Future directions include incorporating multi-modal data, developing semi-supervised learning techniques, optimizing resource efficiency, and creating diverse datasets through collaborative efforts. Emphasizing explainable AI methods will also ensure trust and alignment with clinical reasoning, facilitating the transition of AI models into clinical practice. This work demonstrates the promise of deep learning in CNS tumor classification and highlights key avenues for advancing its clinical utility.
