Adaptive Multimodal Fusion for Histopathological Tumor Analysis : A Hybrid Classical–Quantum Machine Learning Framework

Md Jobair Hossain Faruk

Adaptive Multimodal Fusion for Histopathological Tumor Analysis : A Hybrid Classical–Quantum Machine Learning Framework

2026

Old Westbury, N.Y. : New York Institute of Technology, [2026]

Multimodal user interfaces (Computer systems)
Quantum computing
Diagnostic imaging—Digital techniques
Tumors—Diagnosis—Data processing
Pathology, Cellular—Data processing

English

Accurate analysis of histopathological images is essential for cancer diagnosis, prognostic assessment, and biomarker discovery. However, robust tumor analysis from whole-slide images (WSIs) remains challenging due not only to the gigapixel scale and morphological heterogeneity of histopathological data, but also to limitations of existing AI approaches. Current deep learning models often rely on a single-model paradigm, limiting their ability to simultaneously capture local cellular morphology, global tissue context, and uncertainty associated with complex pathological patterns while maintaining more consistent performance under heterogeneous clinical conditions. This dissertation presents a reinforcement learning (RL)–guided adaptive multimodal fusion framework that integrates complementary computational paradigms for tumor classification and biomarker prediction from histopathological images. The framework consists of four interconnected modules. First, a transformer-based WSI framework for predicting MYCN amplification status directly from neuroblastoma histopathology, demonstrating the feasibility of inferring clinically relevant molecular biomarkers from tissue morphology. Second, a Vision Transformer–based cellular classification framework to distinguish various histopathological cellular classes, enabling fine-grained characterization of tumor morphology at the cellular level. Third, a hybrid classical–quantum machine learning framework that combines transformer-derived image representations with parameterized quantum circuits to investigate quantum-native optimization strategies and computational behavior for high-dimensional pathology data. Finally, an adaptive RL–guided fusion mechanism that dynamically integrates heterogeneous predictive experts, including convolutional, transformer-based, and quantum models, through confidence-aware decision policies. We evaluated the framework using a private multi-institutional neuroblastoma dataset and the publicly available iiiBreakHis dataset. Experimental results demonstrated the effectiveness of the framework for both molecular biomarker prediction and fine-grained morphological analysis. The transformer-based MYCN amplification prediction model achieved 91.8% accuracy. The adaptive fusion framework achieved 98.8% accuracy and consistently outperformed individual expert models and conventional static fusion approaches. In addition, the hybrid classical–quantum framework exhibited distinct computational behavior, demonstrating a slower increase in runtime than the classical baseline within the evaluated experimental setting as the number of whole-slide images increased. Together, these findings demonstrate that RL–guided adaptive fusion of transformerbased, convolutional, and quantum learning models enables more accurate and robust histopathological tumor classification and biomarker prediction than individual models or conventional static fusion approaches. This dissertation contributes a unified computational framework for more stable histopathological tumor analysis and demonstrates how adaptive integration of complementary learning paradigms can support biomarker prediction and tumor classification from digital pathology images.

PDF

College of Engineering & Computing Science

Department of Computer Science

Doctor of Philosophy (Ph.D.) in Computer Science