BiT-Based Early Gastric Cancer Classification Using Endoscopic Images

Hongxiao Li; Shu Li; Xiafei Shi; Xiaoxi Dong; Ge Jin; Lanping Zhu; Yingxin Li; Huijuan Yin

doi:10.3788/LOP202259.0617028

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 6 >Page 0617028 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 6, 0617028 (2022)

BiT-Based Early Gastric Cancer Classification Using Endoscopic Images

Hongxiao Li¹, Shu Li², Xiafei Shi¹, Xiaoxi Dong¹, Ge Jin², Lanping Zhu², Yingxin Li¹, and Huijuan Yin^1、*

Author Affiliations

¹Institute of Biomedical Engineering, Chinese Academy of Medical Sciences & Peking Union Medical College, Tianjin 300192, China

²Department of Gastroenterology, General Hospital of Tianjin Medical University, Tianjin 300050, China

show less

DOI: 10.3788/LOP202259.0617028 Cite this Article Set citation alerts

Hongxiao Li, Shu Li, Xiafei Shi, Xiaoxi Dong, Ge Jin, Lanping Zhu, Yingxin Li, Huijuan Yin. BiT-Based Early Gastric Cancer Classification Using Endoscopic Images[J]. Laser & Optoelectronics Progress, 2022, 59(6): 0617028 Copy Citation Text

show less

Comparison of four metrics among the models with different backbone networks, the spot size represents the number of trainable parameters of each model. (a) Accuracy; (b) F1-score; (c) sensitivity; (d) specificity

Fig. 1. Comparison of four metrics among the models with different backbone networks, the spot size represents the number of trainable parameters of each model. (a) Accuracy; (b) F1-score; (c)

s e n s i t i v i t y

; (d)

s p e c i f i c i t y

Download full size

ROC curves and AUC of different models. (a) ROC curves of five models with different untrainable backbone networks; (b) ROC curves of four models with different trainable backbone networks; (c) ROC curves of six models with a trainable 50×1 backbone network under different batchsizes

Fig. 2. ROC curves and AUC of different models. (a) ROC curves of five models with different untrainable backbone networks; (b) ROC curves of four models with different trainable backbone networks; (c) ROC curves of six models with a trainable 50×1 backbone network under different batchsizes

Download full size

Fig. 3. Confusion matrices of all the models applied on the test set, C represents the cancer label, NC represents the non-cancer label

Download full size

Fig. 4. Examples of testing images

Download full size

Parameter	50×1	101×1	50×3	101×3	152×4
Batchsize	32	16	8	4	4*
Number of parameters	23504450	42496578	211186370	381802178	928356610

Table 1. Batchsizes and parameter amounts of the five BiT backbone networks

Size of backbone network	Trainable state	Accuracy /%	Sensitivity /%	Specificity /%	F1-score /%
50×1	TRUE	95.04	89.33	97.33	91.16
50×1	FALSE	92.37	84.67	95.45	86.39
101×1	TRUE	94.47	86.67	97.59	89.97
101×1	FALSE	94.08	88.67	96.26	89.56
50×3	TRUE	95.99	93.33	97.06	93.02
50×3	FALSE	93.13	83.33	97.06	87.41
101×3	TRUE	97.14	90.67	99.73	94.77
101×3	FALSE	94.85	88.00	97.59	90.72
152×4	FALSE	94.08	82.67	98.66	88.89

Table 2. Testing metrics of the models with different backbone networks on test set

Batchsize	Accuracy /%	Sensitivity /%	Specificity /%	F1-score /%
32	95.04	89.33	97.33	91.16
16	95.23	89.33	97.59	91.47
8	95.23	90.00	97.33	91.53
4	93.89	89.33	95.72	89.33
2	95.23	88.67	97.86	91.41
1	95.42	90.00	97.59	91.84

Table 3. Four metrics of the six models with a trainable 50×1 backbone network under different batchsizes

Parameter	Accuracy	Sensitivity	Specificity	F1-score
Correlation score	0.0835	-0.0784	0.1056	0.0753
p-value	0.8750	0.8826	0.8423	0.8873

Table 4. The Pearson correlation scores and their p-values between the four metrics and the batchsize in Table 3

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information