[论文解读] Using Convolutional Neural Networks for Determining Reticulocyte Percentage in Cats

Krunoslav Vinicki, Pierluigi Ferrari+2 more|arXiv (Cornell University)|Mar 13, 2018

Cell Image Analysis TechniquesBiochemistry, Genetics and Molecular Biology参考文献 6被引用 19

一句话总结

本文提出一种基于深度学习的方法，使用Keras中的预训练单次多框检测器（SSD）自动计数猫血涂片图像中的网织红细胞。该模型仅在800张标注图像上进行训练，预测网织红细胞数量的准确率高达98.7%，表明即使在数据有限的兽医医学领域，也可行实现高精度、易获取的人工智能解决方案。

ABSTRACT

Recent advances in artificial intelligence (AI), specifically in computer vision (CV) and deep learning (DL), have created opportunities for novel systems in many fields. In the last few years, deep learning applications have demonstrated impressive results not only in fields such as autonomous driving and robotics, but also in the field of medicine, where they have, in some cases, even exceeded human-level performance. However, despite the huge potential, adoption of deep learning-based methods is still slow in many areas, especially in veterinary medicine, where we haven't been able to find any research papers using modern convolutional neural networks (CNNs) in medical image processing. We believe that using deep learning-based medical imaging can enable more accurate, faster and less expensive diagnoses in veterinary medicine. In order to do so, however, these methods have to be accessible to everyone in this field, not just to computer scientists. To show the potential of this technology, we present results on a real-world task in veterinary medicine that is usually done manually: feline reticulocyte percentage. Using an open source Keras implementation of the Single-Shot MultiBox Detector (SSD) model architecture and training it on only 800 labeled images, we achieve an accuracy of 98.7% at predicting the correct number of aggregate reticulocytes in microscope images of cat blood smears. The main motivation behind this paper is to show not only that deep learning can approach or even exceed human-level performance on a task like this, but also that anyone in the field can implement it, even without a background in computer science.

研究动机与目标

• 开发一种可访问的、基于深度学习的猫血涂片自动网织红细胞计数方法。
• 证明在兽医医学影像中，仅需极少标注数据即可实现高精度结果。
• 通过使用Keras等开源工具，降低非计算机科学背景人员在兽医医学领域应用该技术的门槛。
• 验证深度学习在真实世界兽医诊断任务中可达到或超越人类水平性能。
• 通过证明该方法在计算和数据需求方面要求适中，推动人工智能在兽医医学中的更广泛应用。

提出的方法

• 作者使用Keras中开源实现的单次多框检测器（SSD）架构进行目标检测。
• 该模型在包含网织红细胞的800张猫血涂片显微镜图像数据集上进行了微调。
• 将网织红细胞视为需检测和计数的对象，使用边界框对每个细胞进行定位。
• 训练过程采用迁移学习和数据增强技术，以提升在小样本数据集上的泛化能力。
• 通过保留的测试集上的标准分类准确率指标对模型进行评估。
• 该实现设计为可复现，并对无高级计算机科学专业知识的用户具有可访问性。

实验结果

研究问题

• RQ1 仅使用800张标注图像，深度学习模型能否在猫血涂片网织红细胞计数中实现高准确率？
• RQ2 在数据有限的兽医医学影像任务中，预训练的SSD模型能否被有效微调？
• RQ3 非计算机科学背景的兽医医学人员是否能够使用开源工具实现并应用此类模型？
• RQ4 该模型在网织红细胞计数中的性能是否超过或匹配人类水平？
• RQ5 该方法能否作为临床环境中人工网织红细胞计数的可扩展、低成本替代方案？

主要发现

• 该模型在预测血涂片图像中网织红细胞总数的测试准确率达到98.7%。
• 尽管仅使用800张标注图像进行训练，该系统仍表现出高度可靠性，表明在小样本数据下具备强大的泛化能力。
• 通过使用迁移学习和基于Keras的SSD框架，实现了无需大量计算资源的有效训练。
• 结果表明，深度学习模型可在兽医网织红细胞计数任务中达到人类水平或更优的性能。
• 该实现对非专家用户友好，作者强调通过开源工具和最低技术要求提升可用性。
• 本研究为在兽医诊断中整合人工智能提供了切实可行的路径，且数据和技术门槛均较低。