【使用cnn降低假正率】Pulmonary Nodule Detection in CT Images笔记

论文全文

INTRODUCTION

这篇文章提出了一种针对肺部结节的CAD系统（Computer-Aided Detection），不仅可以提高正确率，而且降低了假正率。
通常一个CAD系统由两部分组成：nodule candidates detection and false positive reduction. 并且False positives are subsequently reduced in a second stage, which determine most of the performance of CAD systems.
这个系统是由几个cnn网络组合而成。

MATERIALS

介绍训练和交叉验证数据集
LIDC-IDRI
ANODE09
DLCST

METHODS

Candidates Detection

determines the maximum detection sensitivity of subsequent stages.
因为结节形态和大小的差异，将其分成三种：solid nodules,subsolid nodules，Large solid nodules。不同的类有不同的算法，同一个目标三个算法都送入，最后合并三类的结果。

solid nodules

implemented the technique proposed by Murphy et al
具体来说，计算每个三维像素的shape index and curvedness，找出种子点，再将距离近的种子点合并成簇，最后丢弃小的簇。

subsolid nodules

implemented the technique proposed by Jacobs et al
key words:
density mask mask
Morphological opening 先腐蚀后膨胀的过程称为开运算。用来消除小物体、在纤细点处分离物体、平滑较大物体的边界的同时并不明显改变其面积

Large solid nodules

前面两种算法在这种肿瘤上表现不好。另外，由于它胸膜的对比度低，附着在胸膜上的大结节可以用lung segmentation algorithms来排除。we implemented a third detector that consists of three steps.

由于数据的不平衡，分类器可能会倾向常见的组织（比如血管）而忽略肿瘤。为防止过拟合，我们丢弃可能性低的。阈值根据经验设定。

Patches Extraction

从50* 50 * 50mm的立方体中提取图一(b)中的六个特定视角的二维图像
For each candidate，50*50mm，64 64 px；分辨率：0.78 mm，同时对灰度值也有要求。

False Positive Reduction: 2-D Convnets Configuration

The false positive reduction stage is constructed by combining various streams of ConvNets
卷积神经网络的number of views和fusion method在LIDC-IDRI上学习得到，结构是连续的三个卷积层加上一个全连接层。网络结构参见原文图2.
CNN介绍

False Positive Reduction: Convnets Fusion

这里作者调查了三种组合卷积神经网络的方式。注意Mixed-Fusion方法的输入必须是一组正交平面。

Training

888 LIDC-IDRI cases

Data Augmentation

1. Training Data Augmentation
   改变训练图像的大小以及对正样本过采样。
2. Test-Data Augmentation
   Test-data augmentation (TDA) is performed on each candidate (both nodule and non-nodule classes) by rescaling the patches to
   40, 45, 50, and 55 mm

Evaluation

AUC和CMP（在不同误报次数下的召回率情况）

EXPERIMENTAL RESULTS

这一节介绍的是实验结果。
1. fusing models (late-fusion) 的结果比较好。
2. system trained with both contrast and non-contrast scans always
achieves better performance
3. performance of ConvNets improvement when more
views are considered in the architecture

DISCUSSION

这篇文章提出的CAD系统的创新之处在于，利用CNN避免了在降低假正率阶段的人工提取特征和分类。
1. 表1表明combining multiple candidate detection algorithms boosts the detection sensitivity from 85.7% to 93.3% while maintaining a similar number of false positives.
2. Fig. 3(c) shows that incorporating more views in the architecture allows the network to achieve better performance
3. combiner together with other parts of the network gives the best performance
进一步可以：
1. Improvement of the candidate detection algorithm can substantially increase the overall performance of CAD systems, which is planned as future work
2. At 4 FPs/scan, most of undetected nodules are subsolid nodules or nodules with irregular shape.Further data balancing on nodule categories is expected to significantly improve the performance
3. Information from 3-D input data could be exploited to train the ConvNets, even though this would increase the network complexity.
4. Another interesting direction that might also improve performance is by adding features that could not be extracted from patches