Integrating lung parenchyma segmentation and nodule detection with deep multi-task learning
标题: Integrating lung parenchyma segmentation and nodule detection with deep multi-task learning
参考论文:同标题
1. 3D-Unet
2. PDCB(Pyramid dilated convolution block)
选取4种dilation rates,每种dilation rates对应一系列的卷积
每一条卷积操作中,进行类似pyramid convolution(详情见:医学图像处理/2020/11/)的操作,不过卷积核大小是“一样”的,通过嵌入dilation,使得卷积核的扫描范围增大,等效于增大了卷积核。
为了使得不同系列卷积操作最后输出的图像size是一致的,需要将padding size设置为floor(k’/2), k‘表示卷积核实际扫描范围,这也就是Half Padding。
3. 实验细节
- 数据集:The Lung Nodule Analysis 2016(LUNA 16) dataset
- Ten-fold cross-validation experiments :In each round of experiments, we take ninety per- cent of the scans as the training set and the remaining scans as the test set
- Evaluate performance for nodule detection: Free-response receiver operating characteristic(FROC)
- Evaluate performance for lung segmentation: 3D Dice similarity coefficient(DSC)
- 3D depthwise separable
- The size of the feature map from the feature extraction module is set to, W/4 X H/4 X D/4
- encoder: 5 convolution block; decoder: 3 convolution block
- weight decay: 1e-4; momentum: 0.9; batch size: 2
论文中的一些小错误:
- 在PBCD最后一层identity过程中,应该就使用C=128的卷积核,这样才能继续点积,同时作为后续Detection部分输入时,Channel才对的上。 卷积核的Channel数有问题。