Application of a Convolutional Neural Network to Distinguish Burkitt Lymphoma From Diffuse Large B-Cell Lymphoma
Application of a Convolutional Neural Network to Distinguish Burkitt Lymphoma From Diffuse Large B-Cell Lymphoma
复制标题
应用卷积神经网络区分伯基特淋巴瘤和弥漫性大 B 细胞淋巴瘤
DOI：
10.1093/ajcp/aqy099.286
复制
发表时间：
2018
影响因子：
3.5
通讯作者：
Salama, Mohamed
中科院分区：
文献类型：
--
作者：
Mohlman, Jeffrey;Leventhal, Samuel;Venkat, Aniketh;Gyulassy, Attila;Pascucci, Valerio;Salama, Mohamed
伯基特淋巴瘤（BL）和弥漫性大b细胞淋巴瘤（DLBCL）是由于重叠的形态学特征而给诊断带来挑战的实体，通常需要详尽的表型和昂贵的辅助检测才能得出最终诊断。有时，即使进行了广泛的检测，根据细胞遗传学或分子学的发现，最终的诊断可能具有挑战性。卷积神经网络（cnn）是一种流行的物体识别机器学习方法。本研究的目的是评估CNN是否能够可靠地区分BL和DLBCL图像。方法使用Aperio ImageScope （Leica Biosystems, Buffalo Grove， IL）采集单个BL病例的200张图像（×20）和单个DLBCL病例的200张图像（×20）。基于先前发表的工作，开发了一个深度和密集连接的CNN (DenseNet)，并在图像上进行了训练，并在两个保留子集上进行了测试，一个是DLBCL图像（n = 10），另一个是来自原始淋巴瘤病例的10个随机选择的（n = 5 DLBCL和n = 5 BL）未知图像。通过优化小批量训练过程中的交叉熵损失，构建了121层网络。在最后的完全连接层的输出上应用了一个单元型的s型非线性函数。结果输出是图像中每个淋巴瘤类别的预测概率。结果CNN对DLBCL图像子集（10/10）和淋巴瘤未知图像子集（DLBCL = 5/5; BL = 5/5）的预测准确率均为100%，概率均为bb0 99.9%。结论scnn在淋巴瘤亚型的视觉识别中具有应用前景；然而，需要更多的研究来证明这一概念，并且需要一个高效的过程来生成大量高质量的带注释的幻灯片放大图像，以充分测试这一概念，并将cnn应用于更细微的淋巴瘤病例和更广泛的范围。
IntroductionBurkitt lymphoma (BL) and diffuse large B-cell lymphoma (DLBCL) are entities that can present a diagnostic challenge due to overlapping morphological features and often require exhaustive phenotypic and often expensive ancillary testing to yield a final diagnosis. On occasion, even with extensive testing and depending on the cytogenetic or molecular findings, the final diagnosis can be challenging. Convolutional neural networks (CNNs) are a popular machine-learning method for object recognition. The objective of this study was to evaluate if a CNN could reliably differentiate between images of BL and DLBCL.MethodsTwo hundred images (×20) from a single BL case and 200 images (×20) from a single DLBCL case were captured using Aperio ImageScope (Leica Biosystems, Buffalo Grove, IL). Based on previously published work, a deep and densely connected CNN (DenseNet) was developed and trained over the images and tested on two holdout subsets, one of DLBCL images (n = 10) and another of 10 randomly selected (n = 5 DLBCL and n = 5 BL) unknown images from the original lymphoma cases. The 121-layered network was built by optimizing cross-entropy loss during mini-batch training. An element-wise sigmoid nonlinearity function was applied to the outputs of the final, fully connected layer. The resulting output was the predicted probability of each lymphoma class for the image.ResultsThe CNN predicted with 100% accuracy both the subset of DLBCL images (10/10) as well as the lymphoma unknown images (DLBCL = 5/5; BL = 5/5) all with >99.9% probability.ConclusionsCNNs hold promise for visual recognition of lymphoma subtypes; however, more research is needed to prove this concept, and an efficient process for generating extremely large amounts of high-quality magnified images of annotated slides will be necessary to fully test this concept and apply CNNs in more nuanced lymphoma cases and in a broader scope.