Esophageal Squamous Cell Carcinoma Recognition Based on Lightweight Residual Networks with an Attention Mechanism

Jinming Wang; Peng Li; Yan Liang; Wei Sun; Jie Song; Yadong Feng; Lingxiao Zhao

doi:10.3788/LOP220856

Laser & Optoelectronics Progress, Volume. 60, Issue 10, 1010023(2023)

Esophageal Squamous Cell Carcinoma Recognition Based on Lightweight Residual Networks with an Attention Mechanism

Jinming Wang^1,2, Peng Li², Yan Liang³, Wei Sun^1,2, Jie Song³, Yadong Feng³, and Lingxiao Zhao^2、*

¹School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China, Suzhou 215163, Jiangsu, China

²Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Science, Suzhou 215163, Jiangsu, China

³Department of Gastroenterology, Zhongda Hospital, Southeast University, Nanjing 210009, Jiangsu, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(16)

Fig. 1. Architecture of CALite-ResNet

Download full size

Fig. 2. Schematic of GhostModule

Download full size

Fig. 3. Schematic illustration of improved SCConv

Download full size

Fig. 4. Diagrams of residual block structures. (a) Bottleneck structure of ResNet50; (b) GSCBottleneck structure

Download full size

Fig. 5. Structure of CA attention mechanism

Download full size

Fig. 6. Structure of CA-GSCBottleneck module

Download full size

Fig. 7. Schematic of majority voting method

Download full size

Fig. 8. Example diagrams of effects obtained using different data enhancement methods. (a) Origin image; (b) random rotation; (c) horizontal flip; (d) random scaling

Download full size

Fig. 9. ROC curves of different network models

Download full size

Fig. 10. Grad-CAM visualizations

Download full size

Table 1. Classification results of 5-fold cross-validation at image level

View table

Table 1. Classification results of 5-fold cross-validation at image level

Experiment	ACC /%	SENS /%	SPEC /%	PRE /%	F1-score
1-fold	96.30	94.77	97.76	97.59	0.9616
2-fold	94.65	93.95	96.59	98.71	0.9627
3-fold	97.09	97.81	95.63	97.86	0.9783
4-fold	96.82	96.81	96.77	97.78	0.9729
5-fold	97.07	95.14	93.53	94.91	0.9503
Mean±SD	96.39±0.9136	95.70±1.4090	96.06±1.4325	97.37±1.2887	0.9652±0.0097

Table 2. Classification results of 5-fold cross-validation at patient level

View table

Table 2. Classification results of 5-fold cross-validation at patient level

Experiment	ACC /%	SENS /%	SPEC /%	PRE /%	F1-score
1-fold	95.83	94.44	88.89	96.88	0.9564
2-fold	96.15	95.00	90.00	93.33	0.9416
3-fold	95.24	93.75	87.50	96.43	0.9507
4-fold	95.65	95.45	90.90	96.15	0.9580
5-fold	95.65	94.44	88.89	96.67	0.9554
Mean±SD	95.70±0.2953	94.62±0.5731	89.24±1.1494	95.89±1.3039	0.9524±0.0059

Table 3. Comprehensive performance comparison of different network models

View table

Table 3. Comprehensive performance comparison of different network models

Model	Params /10⁶	Predicted time /ms	ACC /%	SENS /%	SPEC /%	F1-score
DenseNet^［27］	12.49	19.40	91.98	91.04	94.57	0.9434
Xception^［28］	20.81	8.31	92.47	91.49	95.19	0.9470
ResNet50^［21］	23.51	11.18	94.32	93.73	95.97	0.9604
ResNeXt50^［29］	22.98	13.61	94.48	93.89	96.12	0.9616
Res2Net50^［30］	23.01	17.05	94.73	94.06	96.59	0.9633
CALite-ResNet	16.66	16.42	94.65	93.95	96.59	0.9627

Table 4. Comparison with related research methods
View table
Table 4. Comparison with related research methods
Method Proposed dataset Open dataset
ACC /% SENS /% F1-score ACC /% SENS /% F1-score
Reference［18］ 94.58 95.60 0.9595 93.43 93.70 0.9415
Proposed method 97.09 97.81 0.9783 96.43 95.62 0.9681

Table 5. Comparison of the results of ablation experiments
View table
Table 5. Comparison of the results of ablation experiments
Method Params /10⁶ ACC /% SENS /% F1-score
Baseline 23.51 96.14 97.21 0.9712
+GSCBottleneck 15.27 95.46 96.80 0.9663
+GSCBottleneck+CA 16.66 96.33 97.33 0.9727
+GSCBottleneck+CA+CBL 16.66 97.09 97.81 0.9783

Table 6. Comparison of experimental results in different attention modules

View table

Table 6. Comparison of experimental results in different attention modules

Method	Params /10⁶	Predicted time /ms	ACC /%	SENS /%	AUC
Baseline（Lite-ResNet）	15.27	16.48	95.41	96.20	0.9715
+ SE	17.27	23.79	95.98	96.61	0.9782
+ CBAM	17.27	21.35	96.69	97.33	0.9826
+ CA（CALite-ResNet）	16.66	16.42	97.09	97.81	0.9883

Tools

Get Citation

Copy Citation Text

Jinming Wang, Peng Li, Yan Liang, Wei Sun, Jie Song, Yadong Feng, Lingxiao Zhao. Esophageal Squamous Cell Carcinoma Recognition Based on Lightweight Residual Networks with an Attention Mechanism[J]. Laser & Optoelectronics Progress, 2023, 60(10): 1010023

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites