Self-Supervised Transfer Learning of Pulmonary Nodule Classification Based on Partially Annotated CT Images

Hong Huang; Chao Peng; Ruoyu Wu; Junli Tao; Jiuquan Zhang

doi:10.3788/AOS202040.1810003

Acta Optica Sinica, Volume. 40, Issue 18, 1810003(2020)

Self-Supervised Transfer Learning of Pulmonary Nodule Classification Based on Partially Annotated CT Images

Hong Huang^1、*, Chao Peng¹, Ruoyu Wu¹, Junli Tao², and Jiuquan Zhang²

¹Key Laboratory of Optoelectronic Technology & Systems, Ministry of Education, Chongqing University, Chongqing 400044, China;

²Department of Radiology, Chongqing University Cancer Hospital, Chongqing 400030, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(9)

Fig. 1. Schematic of self-supervised learning principle

Download full size

Fig. 2. Transformation method of lung CT images. (a) Original image; (b) nonlinear transformation; (c) local pixel change; (d) external pixel change; (e) internal pixel change

Download full size

Fig. 3. Flow chart of proposed algorithm

Download full size

Fig. 4. Loss-accuracy graph of proposed model when FC is 2

Download full size

Fig. 5. ROC curves of different classification models when FC is 2

Download full size

Table 1. Comparison of classification performance under different number of fully connected layers

View table

Table 1. Comparison of classification performance under different number of fully connected layers

Number	Model	A_CC	S_en	S_pe	AUC
	Scratch model	0.815±0.018	0.698±0.065	0.908±0.046	0.879±0.022
FC is 1	Self-supervised model	0.828±0.001	0.684±0.113	0.903±0.043	0.895±0.008
	Proposed model	0.843±0.039	0.754±0.103	0.914±0.020	0.924±0.015
	Scratch model	0.817±0.018	0.751±0.080	0.867±0.054	0.870±0.026
FC is 2	Self-supervised model	0.864±0.013	0.812±0.048	0.903±0.050	0.915±0.009
	Proposed model	0.886±0.009	0.839±0.054	0.920±0.044	0.929±0.016
	Scratch model	0.819±0.048	0.767±0.011	0.861±0.065	0.878±0.039
FC is 3	Self-supervised model	0.830±0.025	0.768±0.068	0.877±0.057	0.905±0.030
	Proposed model	0.841±0.049	0.855±0.030	0.831±0.100	0.906±0.038

Table 2. Comparison of running time of different models under different number of fully connected layerss
View table
Table 2. Comparison of running time of different models under different number of fully connected layerss
Number Scratch model Self-supervised model Proposed model
FC is 1 5.04 4.31 4.34
FC is 2 5.08 4.69 4.68
FC is 3 5.34 4.87 5.39

Table 3. Comparison of classification performance of different algorithms

View table

Table 3. Comparison of classification performance of different algorithms

Algorithm	A_CC	S_en	S_pe	AUC
SVM	0.746±0.032	0.660±0.061	0.814±0.027	0.802±0.025
3D CNN	0.811±0.054	0.740±0.116	0.868±0.040	0.872±0.055
3D ResNet	0.826±0.040	0.758±0.121	0.880±0.050	0.906±0.026
3D DenseNet	0.851±0.028	0.847±0.034	0.854±0.036	0.908±0.010
Proposed algorithm	0.886±0.009	0.839±0.054	0.920±0.044	0.929±0.016

Table 4. Comparison of running time of different algorithms
View table
Table 4. Comparison of running time of different algorithms
Algorithm SVM 3D CNN 3D ResNet 3D DenseNet Proposed algorithm
Time /s 0.08 5.30 7.19 15.23 4.68

Tools

Get Citation

Copy Citation Text

Hong Huang, Chao Peng, Ruoyu Wu, Junli Tao, Jiuquan Zhang. Self-Supervised Transfer Learning of Pulmonary Nodule Classification Based on Partially Annotated CT Images[J]. Acta Optica Sinica, 2020, 40(18): 1810003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Image Processing

Received: Apr. 17, 2020

Accepted: Jun. 11, 2020

Published Online: Aug. 27, 2020

The Author Email: Huang Hong (hhuang@cqu.edu.cn)

DOI:10.3788/AOS202040.1810003

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Comparison of classification performance under different number of fully connected layers

Table 1. Comparison of classification performance under different number of fully connected layers

Table 2. Comparison of running time of different models under different number of fully connected layerss

Table 2. Comparison of running time of different models under different number of fully connected layerss

Table 3. Comparison of classification performance of different algorithms

Table 3. Comparison of classification performance of different algorithms

Table 4. Comparison of running time of different algorithms

Table 4. Comparison of running time of different algorithms