The School of Computer Science is pleased to present…
Hypergraph-Augmented Vision Transformers for Fine-Grained Bone Tumor Subtype Classification
Differentiating primary bone tumor subtypes from X-ray radiographs is a complex task hampered by subtle morphological overlaps and severe class imbalance in real-world datasets. While Convolutional Neural Networks (CNNs) struggle to represent long-range dependencies, standard Vision Transformers (ViTs) are limited to pairwise token interactions, failing to capture the high-order semantic relationships required for fine-grained diagnosis. This thesis proposes two hypergraph-augmented architectures, SoftHGNN-ViT and SoftKHGNN-ViT, which integrate dynamic hypergraph reasoning with a ViT backbone. By utilizing differentiable soft hyperedges and multi-hop message passing via bisection nested convolution, these models effectively capture complex correlations among image regions.
Keywords: Bone Tumor Subtype Classification, Vision Transformers (ViT), Hypergraph Neural Networks, Medical Image Analysis
