| Objective:With the emergence of whole slide images(WSI)and the booming computer vision algorithms,people are increasingly interested in the application of artificial intelligence(AI)on the pathology,especially deep learning(DL)-based AI.At present,researches on the DL-based AI in pathological diagnosis mainly focus on breast cancer,prostate cancer,lung cancer and colorectal cancer and so on,while relatively less in the gastric cancer.This study intends to establish a DL-based AI model assisting the pathological diagnosis of gastric cancer,so as to explore the performance of AI in the pathological diagnosis of primary gastric cancer and lymph node metastasis.Methods:1.Establishing datasets,and setting the ground truth(GT)for different images.1.1 When establishing the dataset of gastric signet ring cell carcinoma,we obtained400× images of GT “cancer” and “non-cancer” from 72 primary gastric signet ring cell carcinoma specimens in our hospital by manually intercepting microscopically.To learn the morphology of signet ring cell more detailed,the model used 224×224 pixel image patch to process the dataset of gastric signet ring cell carcinoma.1.2 When establishing the dataset of multi-subtype gastric cancer,we obtained 400×images of GT “cancer” and “non-cancer” from 125 primary multi-subtype gastric cancer specimens in our hospital by firstly obtaining WSI through digitally scanning and secondly intercepting images.In addition,to observe the adaptability of model to different tissue processing methods,we further obtained400× images of GT “cancer” and “non-cancer” from 98 multi-subtype gastric cancer WSIs in The Cancer Genome Atlas(TCGA)frozen image dataset of gastric cancer by magnifying first and intercepting secondly.To cover more tissue component,the model used 512×512 pixel image patch to process dataset of multi-subtype gastric cancer.1.3 When establishing the dataset of lymph nodes of gastric cancer,we obtained 400×images from non-metastatic lymph nodes and 100× images from metastatic lymph nodes by direct interception microscopically and intercepting followed by scanning respectively,GT is “non-metastasis” and “metastasis” respectively.To cover more tissue component,the model used 512×512 pixel image patch to process dataset of lymph node of gastric cancer.1.4 It is worth noting that,to enhance the generalization ability of model identifying tissues,we added non-metastatic lymph node images of gastric cancer into the dataset of multi-subtype gastric cancer,setting the GT as “non-cancer”.Meanwhile,added cancer image and non-cancer image of primary focus of muti-subtype gastric cancer(our hospital and TCGA)into the dataset of lymph nodes of gastric cancer,setting the GT as “non-metastasis”.2.Distributing training set and test set.Some images were randomly selected from each dataset as training set for the respective dataset,and the remaining images were used as the test set of respective dataset.The training set was used to learn the features of images with different GT,and the test set was used to evaluate the diagnostic performance of DL model.3.Image preprocessing.According to the different image environments and different learning requirements in different datasets,determined the size of image patch to be processed once by the DL model.4.Training and testing model.First,an AI model based on convolutional neural network was established using image from the signet ring cell carcinoma training set,and then the model was tested in the signet ring cell carcinoma test set.Secondly,the data of the multi-subtype gastric cancer training set were added into the model to continue the training,and then be tested in the multi-subtype gastric cancer test set.Finally,lymph node metastasis/non-metastasis of gastric cancer was trained and tested.5.Statistical analysis.SPSS 17.0 software was used for statistical analysis.Index such as accuracy,sensitivity,specificity,positive predictive value,negative predictive value,Youden index,F1-score,Kappa coefficient and AUC of model diagnosis were evaluated by comparing the model-predicting value and GT of each image in test set.Results:1.The distribution of images with different GT1.1 The dataset of gastric signet ring cell carcinoma(our hospital,72 cases,400×,totally 1554 images): including 933 cancerous images(GT “cancer”)and 621 noncancerous images(GT “non-cancer”).1.2 The dataset of multi-subtype gastric cancer(our hospital and TCGA,400×,totally4329 images):(1)125 cases of multi-subtype gastric cancer from our hospital,including 1225 cancerous images(GT “cancer”)and 958 noncancerous images(GT “non-cancer”).(2)98 cases of multi-subtype gastric cancer from TCGA,including 793 cancerous images(GT “cancer”)and 637 noncancerous images(GT “non-cancer”).(3)134 cases of non-metastatic lymph node from our hospital,716 images(GT “non-cancer”).1.3 The dataset of lymph node of gastric cancer(our hospital and TCGA,totally 4991images):(1)61 cases of metastatic lymph node of gastric cancer from our hospital,662 images(GT “metastasis”,100×).(2)134 cases of non-metastatic lymph node of gastric cancer,716 images(GT “non-metastasis”,400×).(3)3613cancerous and non-cancerous images of primary focus of multi-subtype gastric cancer from our hospital and TCGA(GT “non-metastasis”,400×).2.Data distribution of training set and test set.2.1 In the dataset of gastric signet ring cell carcinoma,the training set and test set of GT “cancer” contains 795 images(85.2%)and 138 images(14.8%)respectively,and the training set and test set of GT “non-cancer” contains 498 images(80.2%)and 123 images(19.8%)respectively.2.2 In the dataset of multi-subtype gastric cancer,the training set and test set of GT“cancer” contains 2002 images(99.2%)and 16 images(0.8%)respectively,and the training set and test set of GT “non-cancer” contains 2263 images(97.9%)and 48 images(2.1%)respectively.2.3 In the dataset of lymph node of gastric cancer,the training set and test set of GT“metastasis” contains 646 images(97.6%)and 16 images(2.4%)respectively,and the training set and test set of GT “non-metastasis” contains 4265 images(98.5%)and 64 images(1.5%)respectively.3.Exploration of DL model diagnosing primary focus of gastric cancer3.1 In this study,the model predicted cancer/non-cancer on the patch level in the dataset of gastric signet ring cell carcinoma with accuracy,sensitivity,specificity and AUC of 97.5%,97%,97.9% and 0.995,the rate of missed diagnosis and misdiagnosis were 3% and 2.1% respectively.Through analyzing missed and misdiagnosed cases,the model should learn the unique and diversified morphology of signet ring cell carcinoma further and distinguish with non-neoplastic cells like vascular endothelial cells,plasma cells,tissue cells and fibroblasts.3.2 The model predicted cancer/non-cancer on the image level in the dataset of multi-subtype gastric cancer with the accuracy,sensitivity,specificity and AUC of 90.6%,93.3%,89.8% and 0.943,the rate of missed diagnosis and misdiagnosis were 6.7% and 10.2% respectively.Missed diagnostic cases hinted that the model should reinforce the training in poorly cohesive carcinoma with slight nuclear atypia and be careful to the accuracy of annotation.In case of misdiagnosis,the model should further be trained to recognize non-cancerous epithelial and mesenchymal tissues.4.Exploration of DL model diagnosing lymph node metastasis of gastric cancer4.1 In this study,the model predicted metastasis/non-metastasis on the image level in the dataset of lymph node of gastric cancer with accuracy,sensitivity,specificity and AUC of 83.8%,100%,79.7% and 0.964,the rate of missed diagnosis and misdiagnosis were 0% and 20.3% respectively.Although the model has a good sensitivity(100%)and can identify isolated tumor cells,the test set is small(only16 images of lymph node metastasis),it’s necessary to expand the sample size for further evaluation.4.2 For the misdiagnosed cases,we also need to enhance the capacity of model to distinguish metastatic cancer(especially poorly differentiated cancer)from non-neoplastic cells such as tissue cells,vascular endothelial cells and so on.In addition,there is also a part of false positive cases(6/13)that are primary focus images containing both cancerous tissue and lymphatic tissue,which may be similar to the image pattern of metastatic lymph node,thus leading to misdiagnosis.From another aspect,it also shows how model recognizing lymph node metastasis,that cancer component and lymphoid tissue arise together.Conclusion:1.This study established a DL-based AI model focusing on the special subtype gastric signet ring cell carcinoma prone to missed diagnosis and misdiagnosis clinically.The model can also be applied to multi-subtype gastric cancer,and accommodate not only formalin-fixed paraffin-embedded sections,but also in frozen tissue sections,which has a wide clinical prospect.In addition,the model can also assist pathological diagnosis of gastric cancer lymph node metastasis,which confirm the compatibility in primary and metastatic focus and be convenient for clinical application.2.In the future,we will further train the model in images with larger sample size,more sample types and different magnification,and enhance the identification of non-neoplastic cells and tissues by improving data annotation,in addition,debug the model in multi-center and prospective studies. |
PDF Full Text Request