Seeing the Journey: Progressive Fusion and Prototype Alignment for Generalizable Head-and-Neck Cancer Prognosis

Abstract

Head-and-neck cancer (HNC) prognosis is clinically important for guiding treatment intensity, follow-up scheduling, and supportive care. However, outcome prediction remains chal- lenging due to substantial inter-patient heterogeneity: patients with similar tumour–node– metastasis (TNM) stage and comparable treatment can experience markedly different sur- vival trajectories. This work investigates data-driven survival prediction for HNC using computed tomography (CT) imaging and clinical information, with the goal of improving prognostic accuracy, interpretability, and robustness in realistic clinical settings. A central hypothesis is that (i) prognostic performance benefits from principled multi- modal fusion that reflects the clinical reasoning process, and (ii) deployable models require explicit mechanisms to address multi-center domain shift caused by heterogeneous scanners, acquisition protocols, and documentation practices. To test these hypotheses, the thesis presents two complementary studies. The first is a controlled multimodal fusion benchmark formulated as binary prognosis classification, designed to study how CT imaging and clinical information can be integrated in a clinically meaningful and temporally structured manner. The second extends this setting to censoring-aware survival modeling under multi-center domain shift, focusing on robust representation learning across institutions. First, in a controlled binary prognosis classification setting, the thesis conducts a sys- tematic benchmark of multimodal fusion strategies for CT and clinical metadata, spanning classical concatenation and additive fusion, attention-based approaches (cross-attention, gating, and focal modulation), and representative early- and late-fusion paradigms. Build- ing on the observation that clinicians synthesize patient information sequentially, this work introduces Progressive Multi-modal Fusion (ProMFuse), a clinically grounded ar- chitecture that integrates metadata in stages aligned with the patient journey using gated cross-attention and saliency-based weighting. Under the standardized evaluation setting, ProMFuse demonstrates strong predictive performance and highlights the importance of stage-aware fusion for HNC prognosis. Second, the thesis addresses generalization under institutional heterogeneity by propos- ing UniProto-DSBN (Unified Prototype-aligned Domain-Specific Batch Normalization), a unified multi-center multimodal survival framework trained across two cohorts. UniProto DSBN combines cohort-conditioned normalization via domain-specific batch normalization (DSBN) to absorb low-level cohort-dependent CT appearance statistics with proto type based regularization (compactness and cross-cohort alignment) to encourage cohort- invariant multimodal semantics in a shared prognostic space. Experiments show improved within cohort performance and enhanced cross-cohort robustness compared to naively pooled baselines, supporting unified learning as a step toward practical multi-institution deployment. Overall, this thesis advances multimodal HNC survival modeling by (1) providing com parative evidence across fusion strategies, (2) introducing a clinically coherent progressive fusion mechanism, and (3) proposing a unified multi-center framework that explicitly mitigates domain shift. These contributions aim to improve reliability of prognostic models and facilitate translation to heterogeneous clinical environments.

Related publications

• MedRoundsQA: A Persona and Difficulty Aware Evaluation for Multi-Turn Medical ConsultationsConference on Empirical Methods in Natural Language Processing · 2026 · Same author: Nazira Dunbayeva
• A Machine Learning Approach to Predict Biological Age and Its Longitudinal DriversInternational Joint Conference on Biomedical Engineering Systems and Technologies · 2026 · Same author: Nazira Dunbayeva
• DWX-Net: A Daubechies Wavelet-enhanced Xception Network for Breast Ultrasound Lesion ClassificationIEEE Guwahati Subsection Conference · 2026 · Shared topic: Radiomics and Machine Learning in Medical Imaging
• Convolutional Neural Networks with Transfer Learning for Three-Class Lung CT ClassificationIEEE International Conference on Smart Information Systems and Technologies · 2026 · Shared topic: Radiomics and Machine Learning in Medical Imaging