Segmentation-Driven RetinaNet for Protozoa Detection
Segmentation-Driven RetinaNet for Protozoa Detection
复制标题
DOI：
10.1109/ism.2018.00062
复制
发表时间：
2018-12
期刊：
影响因子：
--
通讯作者：
Khoa Pho;Muhamad Kamal Mohammed Amin;A. Yoshitaka
中科院分区：
文献类型：
--
作者：
Khoa Pho;Muhamad Kamal Mohammed Amin;A. Yoshitaka
原生动物检测和鉴定在寄生虫学、科学研究、生物处理过程和环境质量评价等许多实际领域中发挥着重要作用。传统的原生动物鉴定实验室方法非常耗时，并且需要专业知识和昂贵的设备。另一种方法是使用显微照片来识别原生动物的种类，这样可以节省大量时间并降低成本。然而，这种方法中的现有方法仅在原生动物已经被分割时才能识别物种。这些方法研究形状和尺寸的特征。在这项工作中，我们在图像中检测和识别了各种物种的包囊和卵囊，例如：贾第鞭毛虫、布氏碘阿米巴、刚地弓形虫、卡耶坦环孢子虫、大肠杆菌、肉孢子虫、贝利囊等孢子虫和棘阿米巴，它们的形状共同为圆形，严重影响人类和动物的健康。我们提出分割驱动的 RetinaNet 来自动检测、分割和识别显微照片中的原生动物。通过应用迁移学习、数据增强技术等多种技术，并将训练样本划分为原生动物的生命周期阶段，我们成功克服了深度学习在该问题上应用时缺乏数据的问题。尽管训练数据中每个生命周期类别最多有 5 个样本，但我们提出的方法仍然取得了有希望的结果，并且在我们的原生动物数据集上优于原始 RetinaNet。
Protozoa detection and identification play important roles in many practical domains such as parasitology, scientific research, biological treatment processes, and environmental quality evaluation. Traditional laboratory methods for protozoan identification are time-consuming and require expert knowledge and expensive equipment. Another approach is using micrographs to identify the species of protozoans that can save a lot of time and reduce the cost. However, the existing methods in this approach only identify the species when the protozoan are already segmented. These methods study features of shapes and sizes. In this work, we detect and identify in the images of cysts and oocysts of various species such as: Giardia lamblia, Iodamoeba butschilii, Toxoplasma gondi, Cyclospora cayetanensis, Balantidium coli, Sarcocystis, Cystoisospora belli and Acanthamoeba, which have round shapes in common and affect seriously to human and animal health. We propose Segmentation-driven RetinaNet to automatically detect, segment, and identify protozoans in their micrographs. By applying multiple techniques such as transfer learning, and data augmentation techniques, and dividing training samples into life-cycle stages of protozoans, we successfully overcome the lack of data issue in applying deep learning for this problem. Even though there are at most 5 samples per life-cycle category in the training data, our proposed method still achieves promising results and outperforms the original RetinaNet on our protozoa dataset.