The analysis of leukocyte recruitment in intravital video microscopy (IVM) is essential to the understanding of inflammatory processes. However, because IVM images often present a large variety of visual characteristics, it is hard for an expert human or even conventional machine learning techniques to detect and count the massive amount of cells and extract statistical measures precisely. Convolutional neural networks are a promising approach to overcome this problem, but due to the difficulty of labeling cells, large data sets with ground truth are rare. The present work explores an adaptation of the RetinaNet model with a suite of augmentation techniques and transfer learning for detecting leukocytes in IVM data. The augmentation techniques include simulating the Airy pattern and motion artifacts present in microscopy imaging, followed by traditional photometric, geometric and smooth elastic transformations to reproduce color and shape changes in cells. In addition, we analyzed the use of different network backbones, feature pyramid levels, and image input scales. We have found that even with limited data, our strategy not only enables training without overfitting but also boosts generalization performance. Among several experiments, the model reached a value of 94.84 for the average precision (AP) metric as our best outcome when using data from different image modalities. We also compared our results with conventional image processing techniques and open-source tools. The results showed an outstanding precision of the method compared with other approaches, presenting low error rates for cell counting and centroid distances. Code is available at: https://github.com/brunoggregorio/retinanet-cell-detection.

活体视频显微镜（IVM）中白细胞募集的分析对于理解炎症过程至关重要。但是，由于IVM图像通常表现出各种各样的视觉特征，因此对于专业的人类甚至传统的机器学习技术来说，很难检测和计数大量细胞并精确地提取统计量。卷积神经网络是解决该问题的一种有前途的方法，但是由于标记细胞的困难，具有基本事实的大数据集很少见。本工作探索了一套RetinaNet模型的改编，该模型具有一套增强技术和用于学习IVM数据中白细胞的转移学习。增强技术包括模拟显微镜成像中存在的Airy模式和运动伪影，然后进行传统的光度，几何和平滑弹性转换，以重现细胞的颜色和形状变化。此外，我们分析了不同网络主干网，特征金字塔等级和图像输入比例的使用。我们发现，即使数据有限，我们的策略不仅可以在不过度拟合的情况下进行训练，而且可以提高泛化性能。在一些实验中，当使用来自不同图像模态的数据时，该模型的平均精度（AP）指标达到94.84，这是我们的最佳结果。我们还将我们的结果与传统的图像处理技术和开源工具进行了比较。结果表明，与其他方法相比，该方法具有出色的精度，在细胞计数和质心距离方面具有较低的错误率。代码可在以下网址获得：https：//github.com。