We further experimented with three different CNN architec- tures by training from scratch (i.e., a random set of initial weights). We designed the architectures using Keras29 with Tensorflow30 as the CNN library. The architectures and param- eters used are described in Tables 3– 5. For each locally trained CNN architecture, the input image size was 100 × 100 pixels (used bicubic interpolation for resizing). Different size nodule images (their area varies from 16 to 10,354 pixels) were interpolated to 100 × 100. We performed the experiment with different input image sizes but obtained the best result from 100 × 100. Here, we used the same grayscale CT image for each channel to make it somewhat analogous to an RGB image for the CNN. Doing so engages all the weights and exploits all the knowledge and information during feature extraction from the pretrained network. 解释

这段文字描述了一项实验，其中研究者使用了三种不同的卷积神经网络结构，使用Keras和Tensorflow库进行训练。这些网络结构和参数在表3-5中进行了描述。在每个本地训练的卷积神经网络架构中，输入图像大小为100×100像素。不同大小的结节图像（其面积从16到10,354像素不等）被插值为100×100。尽管使用不同的输入图像大小，但最好的结果是使用100×100大小的图像。使用相同的灰度CT图像作为每个通道的输入，使其在某种程度上类似于CNN中的RGB图像。这样做可以利用预训练网络中的所有权重并从中提取特征。

相关问题

The total number of epochs for training was 200. The learning rate for each architecture was kept constant at 0.0001 with the RMSprop31 (root mean square propagation) algorithm, which was used for gradient descent optimization. Though we experimented with different batch sizes (8/16/24/ 32), a batch size of 16 was used for both training and validat- ing the deep convolutional architecture, because it gave the best result. Leaky ReLU (alpha ¼ 0.01), where negative values are occasionally allowed to propagate, was applied in convolutional layers 1 and 2. This provided nonlinearity on the output of the convolutional layers. As all our architec- tures were shallow, to prevent overfitting, both dropout32 and L2 regularization33 were applied before the classification layer. 解释

这段话主要描述了一个深度卷积神经网络的训练细节。其中，训练的总迭代次数为200次，使用了RMSprop算法作为梯度下降优化器，并设置了学习率为0.0001。在不同的batch size中，选择了16作为训练和验证的批次大小，因为它提供了最好的结果。在前两个卷积层中应用了Leaky ReLU激活函数，这使得负值有时也可以传播，提供了非线性。为了防止过拟合，模型中应用了dropout和L2正则化。

Transfer learning25,26 is a method where previously learned knowledge is applied to another task and the task domains may be different. In our case, the domain is very differ- ent. ImageNet consists of natural camera images and does not include any type of lung nodule or cancer image. Our image set consists of only lung nodules in CT images. We experimented with three different pretrained CNN’s [vgg (visual geometry group)-m/vgg-f/vgg-s]27 in this study using a MA TLAB toolbox named MA TCONVNET.28 The f, m, and s after vgg- stand for fast, medium, and slow and refer to training time (so partly the number of weights). We obtained deep features from the outputs of the last fully connected layer after applying the activation function using a rectified linear unit (post-ReLU), which changes all values <0 to be 0. The LDCT images were grayscale (no color component and we changed the voxel intensities of LDCT images to 0-255 or 8 bit), but the pretrained network was trained on RGB images, so we normalized the images by the average red, green, and blue channel images, and exper- imented by using each channel separately. 解释

Transfer learning是一种方法，其中先前学习的知识被应用于另一个任务，任务域可能不同。在我们的情况下，领域非常不同。ImageNet由自然相机图像组成，不包括任何类型的肺结节或癌症图像。我们的图像集仅包含CT图像中的肺结节。在本研究中，我们使用名为MATCONVNET的MATLAB工具箱尝试了三种不同的预训练CNN（vgg（visual geometry group）-m / vgg-f / vgg-s）进行实验。vgg之后的f，m和s代表快速，中等和慢，并且指的是训练时间（因此部分权重的数量）。我们从应用激活函数后的最后一个完全连接层的输出中获得深层特征，使用修正线性单元（post-ReLU）对所有值<0进行更改以使其为0。LDCT图像是灰度的（没有彩色分量），我们将LDCT图像的体素强度更改为0-255或8位，但预训练网络是在RGB图像上训练的，因此我们通过平均红色，绿色和蓝色通道图像对图像进行归一化，并尝试分别使用每个通道。