结合深度卷积神经网络与影像学特征的肺结节良恶性鉴别方法 
下载： 1337次
Method for Identifying Benign and Malignant Pulmonary Nodules Combing Deep Convolutional Neural Network and Hand-Crafted Features
上海理工大学医疗器械与食品学院, 上海 200082
图 & 表
图 1. 所提方法的流程图
Fig. 1. Flow chart of proposed method
下载图片 查看原文
图 2. 结节融合方法示意图。(a)~(d) 4名放射医师手动分割的肺结节区域;(e)结节融合方法分割的肺结节区域
Fig. 2. Schematic of nodule fusion method. (a)-(d) Pulmonary nodule areas are manually segmented by four radiologists; (e) pulmonary nodule area segmented by nodule fusion method
下载图片 查看原文
图 3. 3D-Inception-ResNet模型架构
Fig. 3. Architecture of 3D-Inception-ResNet model
下载图片 查看原文
图 4. Inception-ResNet模块
Fig. 4. Inception-ResNet module
下载图片 查看原文
图 5. 特征集可视化图
Fig. 5. Visualization map of features
下载图片 查看原文
图 6. LIDC-IDRI数据库中4个典型结节。(a)~(d)肺结节二维切片;(e)~(h)对应结节的三维显示图
Fig. 6. Four typical nodules in LIDC-IDRI database. (a)-(d) 2D slices of nodules; (e)-(h) 3D displays of corresponding nodules
下载图片 查看原文
图 7. 在不同分类器中的ROC曲线。(a) RF; (b) SVM
Fig. 7. ROC curves in different classifiers. (a) RF; (b) SVM
下载图片 查看原文
图 8. 经典CNN架构。(a) 3D-DenseNet模型;(b) 3D-ResNet模型
Fig. 8. Classic CNN architecture. (a)3D-DenseNet model; (b) 3D-ResNet model
下载图片 查看原文
表 1实验设置说明
Table1. Description of experiment setting
| Experiment | Sub experiment | Description |
|---|
| a | Classification based on traditional hand-crafted features | | 1 | b | Classification based on 3D-Inception-ResNet model | | c | Classification combining CNN features and hand-crafted features | | d | Classification on the Shanghai Chest Hospital dataset | | 2 | | Classification under three different sample configuration schemes | | 3 | | Contrast of different architectures (including DenseNet and ResNet) |
|
查看原文
表 2三种肺结节样本配置方案
Table2. Configuration scheme of three pulmonary nodule samples
| Description | Configuration | Number of benignnodules | Number of malignantnodules |
|---|
| ‘1’, ‘2’ as benign and ‘4’, ‘5’ as malignant | 1 | 380 | 300 | | ‘1’, ‘2’, ‘3’ as benign and ‘4’, ‘5’ as malignant | 2 | 736 | 300 | | ‘1’, ‘2’ as benign and ‘3’,‘4’, ‘5’ as malignant | 3 | 380 | 656 |
|
查看原文
表 3实验1分类结果比较
Table3. Comparison of classification results in experiment 1
| Method | A /% | SEN /% | SPE /% | AUC /% |
|---|
| Hand-crafted features+Gaussian-NB | 89.69 | 85.18 | 91.43 | 95.19 | | Hand-crafted features+KNN | 90.72 | 83.33 | 94.03 | 95.54 | | Hand-crafted features+RF | 91.64 | 89.42 | 93.06 | 96.78 | | Hand-crafted features+LDA | 91.75 | 86.21 | 94.12 | 96.18 | | Hand-crafted features+SVM | 91.81 | 88.66 | 94.59 | 96.53 | | 3D-Inception-ResNet | 91.44 | 92.87 | 91.09 | 96.27 | | Combined features+Gaussian-NB | 91.75 | 86.21 | 94.11 | 95.99 | | Combined features+KNN | 91.81 | 88.66 | 94.59 | 96.53 | | Combined features+RF | 92.75 | 92.12 | 93.34 | 97.11 | | Combined features+LDA | 90.72 | 88.46 | 91.55 | 96.46 | | Combined features+SVM | 94.98 | 90.02 | 97.03 | 97.43 |
|
查看原文
表 4SCH数据集分类结果
Table4. Classification results of SCH dataset
| Dataset | A /% | SEN /% | SPE /% | AUC /% |
|---|
| LIDC-IDRI | 94.98 | 90.02 | 97.03 | 97.43 | | SCH | 90.91 | 88.10 | 95.83 | 95.58 |
|
查看原文
表 5三种不同样本配置方案的分类结果比较
Table5. Comparison of classification results in three different sample configuration schemes
| Configuration | Method | A /% | SEN /% | SPE /% | AUC /% |
|---|
| Configuration 1 | Combined features+RF | 92.75 | 92.12 | 93.34 | 97.11 | | Combined features+SVM | 94.89 | 90.02 | 97.03 | 97.43 | | Configuration 2 | Combined features+RF | 86.04 | 70.87 | 90.20 | 91.70 | | Combined features + SVM | 85.86 | 70.20 | 90.55 | 91.12 | | Configuration 3 | Combined features+RF | 80.38 | 85.11 | 72.37 | 87.26 | | Combined features+SVM | 79.73 | 82.35 | 74.91 | 85.31 |
|
查看原文
表 6不同CNN架构的分类结果对比
Table6. Comparison of classification results of different CNN architectures
| Architecture | A /% | SEN /% | SPE /% | AUC /% |
|---|
| 3D-DenseNet | 89.69 | 87.10 | 94.55 | 92.64 | | 3D-DenseNet combining hand-crafted features | 90.72 | 90.00 | 89.47 | 94.20 | | 3D-ResNet | 84.69 | 76.74 | 90.91 | 88.13 | | 3D-ResNet combining hand-crafted features | 90.82 | 90.48 | 91.07 | 95.38 | | 3D-Inception-ResNet | 91.44 | 92.87 | 91.09 | 96.27 | | 3D-Inception-ResNet combining hand-crafted features | 94.98 | 90.02 | 97.03 | 97.43 |
|
查看原文
表 7不同方法的结果比较
Table7. Comparison of the results of different methods
| Method | Number of nodules | A /% | SEN /% | SPE /% | AUC /% |
|---|
| Method in Ref.[26] | 664 | 93.20 | 87.90 | 98.50 | 97.10 | | Method in Ref.[27] | 1226 | 85.62±2.37 | 81.21±6.20 | 89.56±1.17 | 90.45±2.58 | | Proposed method | 1036 | 94.98 | 90.02 | 97.03 | 97.43 |
|
查看原文
高大川, 聂生东. 结合深度卷积神经网络与影像学特征的肺结节良恶性鉴别方法[J]. 光学学报, 2020, 40(24): 2410002. Dachuan Gao, Shengdong Nie. Method for Identifying Benign and Malignant Pulmonary Nodules Combing Deep Convolutional Neural Network and Hand-Crafted Features[J]. Acta Optica Sinica, 2020, 40(24): 2410002.
PDF全文
