Author ORCID Identifier

https://orcid.org/0009-0000-5994-8187

Semester

Summer

Date of Graduation

2026

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Donald Adjeroh

Committee Member

Prashnna Gyawali

Committee Member

Brian Powell

Abstract

Multi-label classification of thoracic diseases from chest X-ray images poses a distinctive challenge: patients rarely present with a single isolated pathology, and automated systems must identify multiple co-occurring conditions simultaneously from a single radiograph. Standard classification architectures treat each disease label as an independent binary decision, failing to capture the statistical co-occurrence patterns, spatial anatomical relationships, and visual feature similarities that characterize multi-disease presentations in clinical data. Class imbalance and incomplete labels compound these problems. In NIH ChestX-ray14, the rarest pathological class has only 227 training instances, and many co-occurring conditions go unlabeled because radiologists focus their reports on the primary finding. This thesis presents GRAFT (Graph-Augmented Framework with Vision Transformers), a multi- stage framework for multi-label chest X-ray classification that integrates self-supervised Vision Transformers with multi-perspective graph-based relational learning. GRAFT operates through three tightly coupled stages: (1) self-supervised pre-training using Masked Autoencoders (MAE), producing domain-specific visual representations without labeled annotations; (2) multi-perspective graph construction using three complementary graph types: a statistical co-occurrence graph with Adaptive Class Balance Correction, a multi-scale spatial anatomical graph using Earth Mover’s Distance at coarse-to-fine resolutions, and a progressive multi-tier visual relationship graph with confidence-based interpolation for sparse disease pairs; and (3) uncertainty-weighted fusion through a Contextual Graph Fusion Network (CGFN) with learnable per-graph uncertainty estimates that dynamically adjust each graph’s contribution. On the NIH ChestX-ray14 benchmark, GRAFT achieved 87.2% mAUC, a 15.0-percentage- point gain over the vision-only baseline (72.2%). The largest improvements appeared in rare and co-occurring conditions: Nodule mAUC improved from 58.3% to 84.8%, and Hernia from 78.5% to 85.5% despite only 227 training instances. On CheXpert, GRAFT achieved 94.0% mAUC, surpassing all prior systems. Given the high prevalence of co-occurring thoracic diseases in aging populations, these results position GRAFT as a practical foundation for multi-disease diagnostic support in elder care.

Download

Share

COinS