Fujita Yoshihisa, Murai Toshiya, Miyata Jun
Department of Psychiatry, Graduate School of Medicine, Kyoto University, Kyoto, Japan.
Department of Psychiatry, Aichi Medical University, Aichi, Japan.
Front Neurosci. 2025 Aug 13;19:1614468. doi: 10.3389/fnins.2025.1614468. eCollection 2025.
Computational saliency map models have facilitated quantitative investigations into how bottom-up visual salience influences attention. Two primary approaches to modeling salience computation exist: one focuses on functional approximation, while the other explores neurobiological implementation. The former provides sufficient performance for applying saliency map models to eye-movement data analysis, whereas the latter offers hypotheses on how neuronal abnormalities affect visual salience. In this study, we propose a novel saliency map model that integrates both approaches. It handles diverse image-derived features, as seen in functional approximation models, while implementing center-surround competition-the core process of salience computation-via an artificial neural network, akin to neurobiological models. We evaluated our model using an open eye-movement dataset and confirmed that its predictive performance is comparable to the conventional saliency map model used in eye-movement analysis. Beyond eye-movement prediction, our model enables neural-level simulations of how neurobiological disturbances influence salience computation. Simulations showed that parameter changes for excitatory-inhibitory balance, baseline neural activity, and synaptic connection density affected the contrast between salient and non-salient objects-in other words-the weighting of salience. Finally, we demonstrated the model's potential for quantifying changes in salience weighting as reflected in eye movements, highlighting its ability to bridge both predictive and neurobiological perspectives. These results present a novel strategy for investigating mechanisms underlying abnormal visual salience.
计算显著性图模型促进了对自下而上的视觉显著性如何影响注意力的定量研究。存在两种主要的显著性计算建模方法：一种侧重于功能近似，另一种则探索神经生物学实现。前者为将显著性图模型应用于眼动数据分析提供了足够的性能，而后者则提供了关于神经元异常如何影响视觉显著性的假设。在本研究中，我们提出了一种整合这两种方法的新型显著性图模型。它处理功能近似模型中所见的各种图像衍生特征，同时通过人工神经网络实现中心-外周竞争（显著性计算的核心过程），类似于神经生物学模型。我们使用一个开放的眼动数据集评估了我们的模型，并确认其预测性能与眼动分析中使用的传统显著性图模型相当。除了眼动预测之外，我们的模型还能够对神经生物学干扰如何影响显著性计算进行神经层面的模拟。模拟表明，兴奋性-抑制性平衡、基线神经活动和突触连接密度的参数变化会影响显著物体与非显著物体之间的对比度，换句话说，就是显著性的权重。最后，我们展示了该模型在量化眼动中反映的显著性权重变化方面的潜力，突出了其在连接预测和神经生物学观点方面的能力。这些结果提出了一种研究异常视觉显著性潜在机制的新策略。
