Table 1.
Details of the datasets used for the development of each AI model.
Table 2.
Definition of each class for training and internal validation.
Fig 1.
Overview of the training of the AI model for daily QC.
An overview of the modeling method used, which includes three steps, is shown. The first step is the data preprocessing step; data are organized and converted in several patches that can be used to train the models. The next step is the patch classifier training step; training is carried out based on the labeled data for each patch generated from a single WSI. The final step is the training WSI classifier step; models that generate information for classification of WSIs are trained by combining the data regarding each WSI. The same model training method was applied to gastric and colorectal models. Abbreviations: AI (artificial intelligence), WSI (whole slide image), QC (quality control), GUI (graphical user interface), M (Malignant), D (Dysplasia), N (Negative for dysplasia).
Fig 2.
The workflow for generating patch images during the data preprocessing step is shown. This workflow involves a series of processes for generating patch images based on the annotation data with WSIs loaded into Openslide’s library, which were processed differently according to class. Abbreviations: WSI (whole slide image).
Fig 3.
Concept of patch classifier training.
The conceptual diagram of the patch image training process is shown.
Fig 4.
Training process for WSI classifiers.
A conceptual diagram of the training process for WSI classifiers is shown. A single WSI is converted into several patch images by the patch maker and distribution information is generated by the patch classifier, which is then converted to a reconstructed WSI. Reconstructed WSIs typically have images rectangular in shape. To classify the reconstructed WSIs, we used a trained CNN model as the WSI classifier. This series of processes was applied to both the gastric and colorectal models to generate two models. Abbreviations: WSI (whole slide image), CNN (convolution neural network).
Fig 5.
Workflow of the operationalization of our trained models.
The operationalization of the trained WSI classifier model is similar to the training workflow. When a new single WSI is loaded into the trained WSI classifier model, the patch maker generates multiple patch images from the WSI, along with patch information, such as the index and location information, which are then recorded in the database. Each patch image is inputted into the patch classifier and the class is inferred by the model. Thus, the classification result is generated and stored in the database. Ultimately, the WSI classifier combines the patch image and database information to generate the reconstructed WSI and classification. This process includes the integration of classification information and location information of each patch for conversion to the reconstructed WSI, as well as input into the WSI classifier for generating WSI classification results and database storage. Therefore, the database stores three types of information: the patch image, patch-level classification, and WSI-label classification information. Patch image information includes the patch index, patch image, and patch location information; patch-level classification information refers to the class inference results generated from the patch classifier model; WSI level classification information refers to the class information inferred by the WSI classifier model for a single WSI. Abbreviations: WSI (whole slide image), M (Malignant), D (Dysplasia), N (Negative for dysplasia).
Fig 6.
Workflow of the SeeDP daily QC system.
In routine practice of gastrointestinal endoscopic biopsy reading, the pathologist first performs the microscopic interpretation. Next, the signed out slides are scanned, then in SeeDP system, their file name (pathology number) is searched for in the database, and the AI model corresponding to each organ is determined and implemented. Finally, via the SeeDP program, scanned slides are organized into classes by pathologic diagnosis and by AI prediction. Pathologists can access the SeeDP system to check whether there is concordance between each class and review discordant cases preferentially. Abbreviations: AI (artificial intelligence), QC (quality control), GUI (graphical user interface), SeeDP (Seegene Medical Foundation’s AI- assisted Digital Pathology Total Solution).
Table 3.
Keywords to determine the appropriate AI model for each biopsy site.
Table 4.
Definition of classes by the pathologic diagnosis for the SeeDP daily QC system.
Table 5.
Keywords for classification by the pathologic diagnosis.
Table 6.
Distribution of the classification by AI prediction of gastric biopsy WSIs.
Table 7.
Distribution of the classification by AI prediction of colorectal biopsy WSIs.
Table 8.
Comparison of detected diagnostic errors and corrections.
Table 9.
Details of seven cases of error identified with the SeeDP daily QC system.