Pan Xue, Xiong Ze, Qiu Dehui, Deng Liguo, Zhao Boxuan, Cao Xiaojie, Wan Xiaohua, Zhang Fa
Department of Ophthalmology, Beijing Shijitan Hospital, Beijing, 100038, China.
Information Engineering College, Capital Normal University, Beijing, 100048, China.
Interdiscip Sci. 2025 Jul 23. doi: 10.1007/s12539-025-00739-x.
Glaucoma is the leading cause of irreversible blindness worldwide. Currently, artificial intelligence (AI) technology combined with optical coherence tomography (OCT) and visual field testing is widely used for glaucoma progression prediction. However, the lack of labeled patient data and the ambiguity of structural differences in glaucoma remain major research bottlenecks. To address this, this paper proposes a glaucoma progression prediction method based on a semi-supervised reinforcement learning (SSRL) framework. In the SSRL framework, we design a key experience filtering strategy (KEFS) that prioritizes high-value training samples to enhance the model's generalization ability and optimize the quality of pseudo-labels. Additionally, an entropy regularization technique is introduced to encourage high-entropy distributions, preventing the model from prematurely converging to specific decisions and improving its ability to explore unlabeled data. To adapt the classifier and the pseudo-label generator within the SSRL framework, we innovatively propose the IncepMambaNet multi-scale network. This network integrates the visual state space (VSS) module from the VMamba model with traditional CNN structures, designing a four-branch IncepMamba module that leverages the strengths of both approaches. Experimental results demonstrate that compared to traditional supervised learning methods, our SSRL model effectively improves the performance of various state-of-the-art (SOTA) supervised classification networks. Furthermore, compared to various SOTA classification networks, IncepMambaNet achieves leading performance in three key metrics-macro-F1, F2-score, and AUC-by 0.8%, 8.8%, and 0.9%, respectively, showcasing outstanding feature capture capability and generalization performance. Ablation studies further confirm that the multi-branch structure and channel reordering operations play a crucial role in enhancing model performance.
青光眼是全球不可逆性失明的主要原因。目前，人工智能（AI）技术与光学相干断层扫描（OCT）和视野测试相结合，被广泛用于青光眼进展预测。然而，缺乏标记的患者数据以及青光眼结构差异的模糊性仍然是主要的研究瓶颈。为了解决这一问题，本文提出了一种基于半监督强化学习（SSRL）框架的青光眼进展预测方法。在SSRL框架中，我们设计了一种关键经验过滤策略（KEFS），对高价值训练样本进行优先排序，以增强模型的泛化能力并优化伪标签的质量。此外，引入了熵正则化技术来鼓励高熵分布，防止模型过早收敛到特定决策，并提高其探索未标记数据的能力。为了使分类器和伪标签生成器在SSRL框架内适配，我们创新性地提出了IncepMambaNet多尺度网络。该网络将VMamba模型中的视觉状态空间（VSS）模块与传统卷积神经网络（CNN）结构相结合，设计了一个四分支的IncepMamba模块，利用了两种方法的优势。实验结果表明，与传统监督学习方法相比，我们的SSRL模型有效地提高了各种先进（SOTA）监督分类网络的性能。此外，与各种SOTA分类网络相比，IncepMambaNet在三个关键指标——宏F1、F2分数和AUC上分别领先0.8%、8.8%和0.9%，展现出卓越的特征捕捉能力和泛化性能。消融研究进一步证实，多分支结构和通道重排操作在增强模型性能方面起着关键作用。
