Enhancing Bone Tumor Classification in X ray Radiographs via Graph Theoretic Priors in Vision Transformers
MSc Thesis Defense by:
Bardiya Rasekh
Date: March 30th
Time: 1 PM
Location: 122 Essex Hall
Abstract:
This thesis addresses the challenges of semantic ambiguity and data scarcity in classifying bone tumor subtypes from X-ray radiographs by integrating graph-theoretic priors into Vision Transformers. The proposed framework introduces novel hybrid architectures that utilize hypergraph aggregation and structural alignment to capture complex interactions and subtle radiographic features. Experimental results demonstrate that these structure-aware models consistently outperform conventional deep learning benchmarks. This research underscores the importance of incorporating high-order structural information into Transformer-based models to achieve the high diagnostic precision required for specialized medical imaging applications.
Keywords: Bone Tumor Classification, Vision Transformers, Graph Neural Networks, Medical Image Analysis, Deep Learning
Thesis Committee:
Internal Reader: Dr. Luis Rueda
External Reader: Dr. Mehdi Sangani Monfared
Advisor: Dr. Alioune Ngom
Chair: Dr. Arunita Jaekel
Registration Link (MAC students Only)
