Application of deep learning in image recognition and diagnosis of gastric cancer

Table 1 Commonly used databases in image recognition of gastric cancer

Database	Time collected	Number of samples	Resolution	Training set	Test set
GR-AIDS[31]	2019	1036496	512 × 512	829197	103650
Jang Hyung Lee[32]	2019	787	224 × 224	717	70
Toshiaki Hirasawa[33]	2018	13584	512 × 512	13584	2496
Bum-Joo Cho[34]	2019	5017	512 × 512	4205	812
Hiroya Ueyama[35]	2020	7874	512 × 512	5574	2300
Lan Li[36]	2020	2088	512 × 512	1747	341
Mads Sylvest Bergholt[37]	2011	1063	512 × 512	850	213

Table 2 Specific concepts of the main evaluation indicators

Index	Description	Usage	Unit
DICE	Repeat rate between the segmentation results and markers	Commonly	%
RMSD	The root mean square of the symmetrical position surface distance between the segmentation results and the markers	Commonly	mm
VOE	The degree of overlap between the segmentation results and the actual segmentation results represents the error rate	Commonly	%
RVD	The difference in volume between the segmentation results and the markers	Rarely	%

DICE: DICE Similarity Coefficient; RMSD: Root-Mean-Square Deviation; VOE: Volumetric Over-lap Error; RVD: Relative Volume Difference.

Table 3 Comparison of recognition performance of convolutional neural network, full convolutional neural network, and ensemble convolutional neural network models

Methods	DICE/%	VOE/%	RMSD/mm
Toshiaki Hirasawa (CNN)	0.5738	0.5977	6.491
Hiroya Ueyama (CNN)	0.6327	0.5373	7.257
Jang Hyung Lee (FCN)	0.8102	0.319	2.468
Bum-Joo Cho (FCN)	0.9350	0.1221	-
Dat Tien Nguyen (ECNN)	0.8947	0.113	-

CNN: Convolutional neural network; FCN: Full convolutional neural network; ECNN: Evolutionary convolutional neural network; DICE: DICE Similarity Coefficient; VOE: Volumetric Over-lap Error; RMSD: Root-Mean-Square Deviation.

Table 4 Comparison of convolutional neural network, full convolutional neural network, and generative adversarial network models

Model features	Contributions	Advantages	Disadvantages	Scope of application
CNN	The topology can be extracted from a two-dimensional image, and the backpropagation algorithm is used to optimize the network structure and solve the unknown parameters in the network	Shared convolution kernel, processing high-dimensional data without pressure; Feature extraction can be done automatically	When the network layer is too deep, the parameters near the input layer will be changed slowly by using BP propagation to modify parameters. A gradient descent algorithm is used to make the training results converge to the local minimum rather than the global minimum. The pooling layer will lose a lot of valuable information	Suitable for data scenarios with similar network structures
FCN	The end-to-end convolutional network is extended to semantic segmentation. The deconvolution layer is used for up-sampling; A skip connection is proposed to improve the roughness of the upper sampling	Can accept any size; Input image; Jump junction; The structure combines fine layers and coarse; Rough layers, generating precise segmentation	The receptive field is too small to obtain the global information;Small storage overhead	Applicable to large sample data
GAN	With adversarial learning criteria, there are two No's: The same network, not a single network	Can produce a clearer, more realistic sample; any generated network can be trained	Training is unstable and difficult to train; GAN is not suitable for processing data in discrete form	Suitable for data generation (e.g., there are not many data sets with labels), image style transfer; Image denoising and restoration; Used to counter attacks

CNN: Convolutional neural network; FCN: Full convolutional neural network; GAN: Generative adversarial network.