Skip to main content
. 2021 Dec;13(12):7006–7020. doi: 10.21037/jtd-21-806

Table 3. The characteristics of the included studies regarding DL models for identifying tumor patterns or variety of cells on digital slides.

Authors Publication year Number of datasets Number of cases Number of images Subtypes (images) Training set (images/cells) Validation set (images/cells) Test set (images/cells) Independent test datasets
(images/cells)
Classifier Results Conclusion
ACC SN SP AUC Precision
Gertych et al. (38) 2019 3 110 206 CSMC: ADC (n=91);
MIMW: ADC (n=88);
TCGA: ADC (n=27)
78 19 109 0 GoogLeNet, ResNet-50 and AlexNet FT-AlexNet:75.3%;
DN-AlexNet-1: 89.90%;
GoogLeNet-2: 85.84%;
Resnet-50-3: 87.64%
NR NR NR NR One of the DN-AlexNets obtained the best performance than other CNNs, with the accuracies of 89.90% for classification involving the five tissue classes in test set
Wei et al. (12) 2019 1 NR 422 DHMC: ADC (n=422) 245 34 143 0 ResNet NR NR NR Lepidic: 0.988;
acinar: 0.970;
papillary: 0.993;
micropapillary: 0.981;
solid: 0.997;
benign: 0.988
NR CNN could improve classification accuracy of ADC patterns by automatically pre-screening, superior to pathologists
Wang et al. (7) 2020 2 507 639 TCGA: ADC (n=208);
NLST: ADC (n=431)
12,000 cell nuclei 1,227 cell nuclei 1,086 cell nuclei 0 Mask R-CNN, Cox proportional hazard prognostic model 88% in the validation set; 90% in the testing set. NR NR NR NR Mask R-CNN extracted and identified 48 cell spatial features, which could predict high-risk group, significantly worse survival than the low-risk group
AbdulJabbar et al. (41) 2020 2 1,070 4,599 TRACERx: NSCLC (n=275);
LATTICe-A: ADC (n=4,324)
16790 H&E cells and 9333 IHC cells 4219 H&E cells 5951 H&E cells and 5028 IHC cells 5082 H&E cells SCCNN Lymphocyte: 0.942;
tumor: 0.933;
other: 0.917;
stromal: 0.936
Lymphocyte: 0.902;
tumor: 0.936;
other: 0.853;
stromal: 0.898
Lymphocyte: 0.982;
tumor: 0.930;
other: 0.981;
stromal: 0.973
NR NR SCCNN for NSCLCs exhibited high accuracy of single-cell classification in H&E digital slides and T-cell identification in the IHC image slides, respectively
Wang et al. (40) 2019 3 NR 159 TCGA and NLST: ADC (n=29);
SPORE: ADC (n=130)
29 130 0 0 DL-based ConvPath software Lymphocytes: 99.3%;
stromal cells: 87.9%;
tumor cells: 91.6%;
overall: 92.9%
NR NR NR NR The overall classification accuracies of the CNN in both datasets were 99.3% for lymphocytes, 87.9% for stromal cells, and 91.6% for tumor cells, respectively
Teramoto
et al. (42)
2020 1 60 793 Normal (n=25);
malignant (n=35)
NR 173 NR 0 PGGAN, DCGAN, ImageNet ImageNet: 0.810;
DCGAN: 0.795;
PGGAN: 0.853
ImageNet: 0.850;
DCGAN: 0.793;
PGGAN: 0.854
ImageNet: 0.768;
DCGAN: 0.797;
PGGAN: 0.853
NR NR PGGAN for cytological specimens improved the classification specificity by 8.5% and the total classification accuracy by approximately 4.3% compared to a CNN model
Saha et al. (43) 2021 1 712 712 TCGA: ADC (n=356);
SCC (n=356)
356 160 160 0 TilGAN 0.98 0.96 NR NR 0.98 TilGAN generated the high quality of synthetic pathology images could efficiently classify real TIL and non-TIL patches with improved accuracy

NR, not reported; CNN, convolutional neural networks; DL, deep learning; SCCNN, sensitive convolutional neural networks; ADC, adenocarcinoma; NSCLC, non-small cell lung cancer; TCGA, The Cancer Genome Atlas; CSMC, Cedars-Sinai Medical Center; MIMW, the Military Institute of Medicine in Warsaw; DHMC, the Dartmouth-Hitchcock Medical Center; NLST, the National Lung Screening Trial project; SPORE, the University of Texas Special Program of Research Excellence; H&E, hematoxylin and eosin; IHC, immunohistochemistry; ACC, accuracy; SN, sensitivity; SP, specificity; GAN, generative adversarial network; PGGAN, progressive growing of GAN; DCGAN, deep convolutional generative adversarial network; TIL, tumor-infiltrating lymphocyte.