Mahmood Tahir, Arsalan Muhammad, Owais Muhammad, Lee Min Beom, Park Kang Ryoung
Division of Electronics and Electrical Engineering, Dongguk University, 30 Pildong-ro 1-gil, Jung-gu, Seoul 04620, Korea.
J Clin Med. 2020 Mar 10;9(3):749. doi: 10.3390/jcm9030749.
Breast cancer is the leading cause of mortality in women. Early diagnosis of breast cancer can reduce the mortality rate. In the diagnosis, the mitotic cell count is an important biomarker for predicting the aggressiveness, prognosis, and grade of breast cancer. In general, pathologists manually examine histopathology images under high-resolution microscopes for the detection of mitotic cells. However, because of the minute differences between the mitotic and normal cells, this process is tiresome, time-consuming, and subjective. To overcome these challenges, artificial-intelligence-based (AI-based) techniques have been developed which automatically detect mitotic cells in the histopathology images. Such AI techniques accelerate the diagnosis and can be used as a second-opinion system for a medical doctor. Previously, conventional image-processing techniques were used for the detection of mitotic cells, which have low accuracy and high computational cost. Therefore, a number of deep-learning techniques that demonstrate outstanding performance and low computational cost were recently developed; however, they still require improvement in terms of accuracy and reliability. Therefore, we present a multistage mitotic-cell-detection method based on Faster region convolutional neural network (Faster R-CNN) and deep CNNs. Two open datasets (international conference on pattern recognition (ICPR) 2012 and ICPR 2014 (MITOS-ATYPIA-14)) of breast cancer histopathology were used in our experiments. The experimental results showed that our method achieves the state-of-the-art results of 0.876 precision, 0.841 recall, and 0.858 F1-measure for the ICPR 2012 dataset, and 0.848 precision, 0.583 recall, and 0.691 F1-measure for the ICPR 2014 dataset, which were higher than those obtained using previous methods. Moreover, we tested the generalization capability of our technique by testing on the tumor proliferation assessment challenge 2016 (TUPAC16) dataset and found that our technique also performs well in a cross-dataset experiment which proved the generalization capability of our proposed technique.
乳腺癌是女性死亡的主要原因。乳腺癌的早期诊断可以降低死亡率。在诊断中，有丝分裂细胞计数是预测乳腺癌侵袭性、预后和分级的重要生物标志物。一般来说，病理学家在高分辨率显微镜下手动检查组织病理学图像以检测有丝分裂细胞。然而，由于有丝分裂细胞与正常细胞之间的细微差异，这个过程既繁琐又耗时，而且具有主观性。为了克服这些挑战，已经开发了基于人工智能（AI）的技术，可自动检测组织病理学图像中的有丝分裂细胞。此类AI技术加快了诊断速度，可作为医生的第二意见系统。以前，传统的图像处理技术用于检测有丝分裂细胞，其准确性低且计算成本高。因此，最近开发了一些表现出色且计算成本低的深度学习技术；然而，它们在准确性和可靠性方面仍需改进。因此，我们提出了一种基于更快区域卷积神经网络（Faster R-CNN）和深度卷积神经网络（CNN）的多阶段有丝分裂细胞检测方法。我们的实验使用了两个乳腺癌组织病理学的开放数据集（2012年国际模式识别会议（ICPR）和2014年ICPR（MITOS-ATYPIA-14））。实验结果表明，对于ICPR 2012数据集，我们的方法实现了0.876的精确率、0.841的召回率和0.858的F1值的最优结果，对于ICPR 2014数据集，实现了0.848的精确率、0.583的召回率和0.691的F1值，高于使用以前方法获得的结果。此外，我们通过在2016年肿瘤增殖评估挑战赛（TUPAC16）数据集上进行测试来检验我们技术的泛化能力，发现我们的技术在跨数据集实验中也表现良好，这证明了我们提出的技术的泛化能力。
