Early identification of human pathogens is crucial for the effective treatment of bloodstream infections to prevent sepsis. Since pathogens that are present in small numbers are usually difficult to detect directly, we hypothesize that the behavior of the immune cells that are present in large numbers may provide indirect evidence about the causative pathogen of the infection. We previously applied time-lapse microscopy to observe that neutrophils isolated from human whole-blood samples, which had been infected with the human-pathogenic fungus or , indeed exhibited a characteristic morphodynamic behavior. Tracking the neutrophil movement and shape dynamics over time, combined with machine learning approach, the accuracy for the differentiation between the two species was about 75%. In this study, the focus is on improving the classification accuracy of the species using advanced deep learning methods. We implemented (i) gated recurrent unit (GRU) networks and transformer-based networks for video data, and (ii) convolutional neural networks (CNNs) for individual frames of the time-lapse microscopy data. While the GRU and transformer-based approaches yielded promising results with 96% and 100% accuracy, respectively, the classification based on videos proved to be very time-consuming and required several hours. In contrast, the CNN model for individual microscopy frames yielded results within minutes, and, utilizing a majority-vote technique, achieved 100% accuracy both in identifying the pathogen-free blood samples and in distinguishing between the species. The applied CNN demonstrates the potential for automatically differentiating bloodstream infections with high accuracy and efficiency. We further analysed the results of the CNN using explainable artificial intelligence (XAI) techniques to understand the critical features and patterns, thereby shedding light on potential key morphodynamic characteristics of neutrophils in response to different species. This approach could provide new insights into host-pathogen interactions and may facilitate the development of rapid, automated diagnostic tools for differentiating fungal species in blood samples.

中文翻译：

---

基于深度学习的离体人念珠菌血液感染中性粒细胞激活表型的表征

早期识别人类病原体对于有效治疗血流感染以预防败血症至关重要。由于少量存在的病原体通常难以直接检测，因此我们假设大量存在的免疫细胞的行为可能提供有关感染病原体的间接证据。我们之前应用延时显微镜观察从感染人类致病真菌的人类全血样本中分离出的中性粒细胞，确实表现出特征性的形态动力学行为。跟踪中性粒细胞随时间的运动和形状动态，结合机器学习方法，区分两个物种的准确度约为 75%。本研究的重点是利用先进的深度学习方法提高物种的分类准确性。我们针对视频数据实现了（i）门控循环单元（GRU）网络和基于变压器的网络，以及（ii）针对延时显微镜数据的各个帧的卷积神经网络（CNN）。虽然 GRU 和基于 Transformer 的方法分别产生了 96% 和 100% 准确率的可喜结果，但基于视频的分类被证明非常耗时，需要几个小时。相比之下，针对单个显微镜帧的 CNN 模型在几分钟内就得出了结果，并且利用多数表决技术，在识别无病原体血液样本和区分物种方面都达到了 100% 的准确性。所应用的 CNN 展示了以高精度和高效率自动区分血流感染的潜力。我们使用可解释的人工智能（XAI）技术进一步分析了 CNN 的结果，以了解关键特征和模式，从而揭示中性粒细胞响应不同物种的潜在关键形态动力学特征。这种方法可以为宿主与病原体相互作用提供新的见解，并可能促进快速、自动化诊断工具的开发，以区分血液样本中的真菌种类。