Leveraging Object Detection for Early Diagnosis of Neurodegenerative Diseases through Radiomic Analysis

Linjian Huang^1,2Miaoran Li³Xin Xu¹Zhijian Xie^4*

• ¹The Second Affiliated Hospital Zhejiang University School of Medicine, Hangzhou, China
• ²Zhejiang University School of Medicine, Hangzhou, China
• ³Hangzhou Dental Hospital West Branch, Hangzhou, China
• ⁴The Key Laboratory of Reproductive Genetics (Zhejiang University), Ministry of Education, Zhejiang university school of medicine, Hangzhou，Zhejiang, China

Early diagnosis of neurodegenerative diseases remains a formidable challenge in modern neuroimaging, due to subtle and heterogeneous brain deterioration patterns in early disease stages. Integrating artificial intelligence and radiomic analysis has emerged as a powerful paradigm for non-invasive biomarker discovery and precision diagnostics. In alignment with trends emphasizing cross-modality analysis, interpretability, and demographic generalization, this study introduces a novel approach leveraging object detection and disentangled representation learning to improve early detection sensitivity and reliability. Traditional radiomics frameworks often suffer from limited generalizability, rigid feature engineering, and confounding variability from age, imaging protocol, or anatomical variations, undermining clinical robustness. Our method addresses these limitations through a three-pronged strategy. We construct a hybrid representation framework separating age-related morphometric changes from disease-specific alterations. We introduce NeuroFact-Net, a dual-path variational encoder-decoder architecture supervised along anatomical and diagnostic axes, enhancing interpretability and facilitating trajectory analysis. We devise a Causal Disease-Aware Alignment (CDAA) strategy imposing population-level invariance and disease-specific consistency using contrastive learning, adversarial subgroup confusion, and maximum mean discrepancy constraints. Experiments across multi-site MRI and PET datasets demonstrate superior diagnostic accuracy, domain transferability, and latent biomarker interpretability, validating its potential for clinical deployment in early-stage screening. This work contributes a scalable, interpretable, and causally grounded computational framework aligned with AI-enhanced neuroimaging advancements.