Abstract

Lymphoma has become the seventh most common cancer expected to occur and the ninth most common cause of cancer death in both males and females. However, pathological diagnosis as the main diagnostic method is time-consuming, expensive and error-prone. Nowadays, with the outstanding performance of Convolutional Neural Network (CNN) in image analysis, its application in medical image classification and segmentation is becoming more and more widespread. Since Faster R-CNN achieves state-of-the-art object detection accuracy on PASCAL VOC 2007, 2012, and MS COCO datasets, in this work, we use it to identify lymphoma cells from a dataset of blood cells. In order to achieve this goal, first, we focus on a new blood cell dataset that mainly consists of lymphoma cells, blasts and lymphocytes. This dataset will be used to fine-tune a pre-trained network. Then, we use two training methods and three networks to train the same dataset respectively. Finally, we choose the best trained model to diagnose lymphoma cells in a new dataset which also contains lymphoma cells, blasts and lymphocytes. The detection rate of lymphoma cells was higher than 96%, and the false detection rate was less than 13%, which is an improvement compared with the previously proposed results. The results show a potential of the proposed method in lymphoma diagnosis.

摘要

在男性和女性中，淋巴瘤已成为第七大最常见的癌症，也是第九大最常见的癌症死亡原因。然而，病理诊断作为主要诊断方法是费时，昂贵且容易出错的。如今，随着卷积神经网络（CNN）在图像分析中的出色表现，其在医学图像分类和分割中的应用越来越广泛。由于Faster R-CNN在PASCAL VOC 2007、2012和MS COCO数据集上均达到了最新的对象检测精度，因此在这项工作中，我们将其用于从血细胞数据集中识别淋巴瘤细胞。为了实现这一目标，首先，我们关注一个主要由淋巴瘤细胞，原始细胞和淋巴细胞组成的新血细胞数据集。该数据集将用于微调预训练的网络。然后，我们使用两种训练方法和三个网络分别训练同一数据集。最后，我们在新的数据集中选择受过最佳训练的模型来诊断淋巴瘤细胞，该数据集还包含淋巴瘤细胞，胚细胞和淋巴细胞。淋巴瘤细胞的检出率高于96％，错误的检出率低于13％，与先前提出的结果相比有所改善。结果显示了该方法在淋巴瘤诊断中的潜力。与之前提出的结果相比，这是一个改进。结果显示了该方法在淋巴瘤诊断中的潜力。与之前提出的结果相比，这是一个改进。结果显示了该方法在淋巴瘤诊断中的潜力。