Ben Hamida A, Devanne M, Weber J, Truntzer C, Derangère V, Ghiringhelli F, Forestier G, Wemmert C
ICube, University of Strasbourg, France.
IRIMAS, University of Haute-Alsace, France.
Comput Biol Med. 2021 Sep;136:104730. doi: 10.1016/j.compbiomed.2021.104730. Epub 2021 Aug 4.
Nowadays, digital pathology plays a major role in the diagnosis and prognosis of tumours. Unfortunately, existing methods remain limited when faced with the high resolution and size of Whole Slide Images (WSIs) coupled with the lack of richly annotated datasets. Regarding the ability of the Deep Learning (DL) methods to cope with the large scale applications, such models seem like an appealing solution for tissue classification and segmentation in histopathological images. This paper focuses on the use of DL architectures to classify and highlight colon cancer regions in a sparsely annotated histopathological data context. First, we review and compare state-of-the-art Convolutional Neural networks (CNN) including the AlexNet, vgg, ResNet, DenseNet and Inception models. To cope with the shortage of rich WSI datasets, we have resorted to the use of transfer learning techniques. This strategy comes with the hallmark of relying on a large size computer vision dataset (ImageNet) to train the network and generate a rich collection of learnt features. The testing and evaluation of such models on our AiCOLO colon cancer dataset ensure accurate patch-level classification results reaching up to 96.98% accuracy rate with ResNet. The CNN models have also been tested and evaluated with the CRC-5000, nct-crc-he-100k and merged datasets. ResNet respectively achieves 96.77%, 99.76% and 99.98% for the three publicly available datasets. Then, we present a pixel-wise segmentation strategy for colon cancer WSIs through the use of both UNet and SegNet models. We introduce a multi-step training strategy as a remedy for the sparse annotation of histopathological images. UNet and SegNet are used and tested in different training scenarios including data augmentation and transfer learning and ensure up to 76.18% and 81.22% accuracy rates. Besides, we test our training strategy and models on the CRC-5000, nct-crc-he-100k and Warwick datasets. Respective accuracy rates of 98.66%, 99.12% and 78.39% were achieved by SegNet. Finally, we analyze the existing models to discover the most suitable network and the most effective training strategy for our colon tumour segmentation case study..
如今，数字病理学在肿瘤的诊断和预后中发挥着重要作用。不幸的是，面对全切片图像（WSIs）的高分辨率和大尺寸，以及缺乏丰富注释的数据集时，现有方法仍然存在局限性。关于深度学习（DL）方法处理大规模应用的能力，此类模型似乎是组织病理学图像中组织分类和分割的一个有吸引力的解决方案。本文重点研究在稀疏注释的组织病理学数据背景下，使用DL架构对结肠癌区域进行分类和突出显示。首先，我们回顾并比较了包括AlexNet、vgg、ResNet、DenseNet和Inception模型在内的当前最先进的卷积神经网络（CNN）。为了应对丰富的WSI数据集短缺的问题，我们采用了迁移学习技术。这种策略的特点是依赖大型计算机视觉数据集（ImageNet）来训练网络并生成丰富的学习特征集合。在我们的AiCOLO结肠癌数据集上对这些模型进行测试和评估，确保了精确的切片级分类结果，ResNet的准确率高达96.98%。CNN模型也在CRC - 5000、nct - crc - he - 100k和合并数据集上进行了测试和评估。对于这三个公开可用的数据集，ResNet分别达到了96.77%、99.76%和99.98%。然后，我们通过使用UNet和SegNet模型，提出了一种针对结肠癌WSIs的逐像素分割策略。我们引入了一种多步骤训练策略，作为对组织病理学图像稀疏注释的一种补救措施。UNet和SegNet在包括数据增强和迁移学习在内的不同训练场景中使用和测试，并确保了高达76.18%和81.22%的准确率。此外，我们在CRC - 5000、nct - crc - he - 100k和沃里克数据集上测试了我们的训练策略和模型。SegNet分别实现了98.66%、99.12%和78.39%的准确率。最后，我们分析现有模型，以发现适合我们结肠癌分割案例研究的最合适网络和最有效的训练策略。
