人工智能
计算机科学
代表(政治)
特征(语言学)
特征学习
计算机视觉
自然语言处理
边距(机器学习)
模式识别(心理学)
光学(聚焦)
作者
Wanyu Zhang,Chengxin Zheng,Xiaodan Zhang,Jianing Wang,Yue Zhang,Xiaobing Yu
出处
期刊:Multidimensional Signal Processing Workshop
[Institute of Electrical and Electronics Engineers]
日期:2026-06-01
卷期号:: 1-6
标识
DOI:10.1109/ivmsp68335.2026.11618830
摘要
Learning a compact and nuisance-robust longitudinal representation from paired OCT scans is essential for reliable retinal prognosis. However, follow-up acquisitions often vary in device settings, speckle statistics, and mild inter-visit misregistration, making it difficult for existing deep models to disentangle true pathological progression from acquisition-induced perturbations. To address this challenge, we propose a diffusion-driven longitudinal representation learning framework for prognosis classification from paired OCT images. The core design is a two-timepoint differential representation module that explicitly models treatment-induced changes by contrasting refined latent representations of pre- and post-treatment scans. We further adopt a parameter-efficient strategy that combines the latent representation space of Stable Diffusion with a LoRA-adapted LLaMA backbone used as a generic sequence modeling module over latent change tokens, enabling endpoint-specific decoding without full-model fine-tuning. In addition, we introduce targeted speckle-noise and affine-jitter perturbations, as well as a twoview consistency regularizer, encouraging prediction invariance under clinically plausible disturbances. Experiments demonstrate that our method achieves improved overall performance on the evaluated dataset for dual-outcome prognosis classification, suggesting its potential utility for longitudinal OCT prognosis. The code will be made publicly available upon acceptance.
科研通智能强力驱动
Strongly Powered by AbleSci AI