Artificial intelligence (AI)-based analysis has recently been adopted in the examination of histological slides via the digitization of glass slides using a digital scanner. In this study, we examined the effect of varying the staining color tone and magnification level of a dataset on the result of AI model prediction in hematoxylin and eosin stained whole slide images (WSIs). The WSIs of liver tissues with fibrosis were used as an example, and three different datasets (N20, B20, and B10) were prepared with different color tones and magnifications. Using these datasets, we built five models trained Mask R-CNN algorithm by a single or mixed dataset of N20, B20, and B10. We evaluated their model performance using the test dataset of three datasets. It was found that the models that were trained with mixed datasets (models B20/N20 and B10/B20), which consist of different color tones or magnifications, performed better than the single dataset trained models. Consequently, superior performance of the mixed models was obtained from the actual prediction results of the test images. We suggest that training the algorithm with various staining color tones and multi-scaled image datasets would be more optimized for consistent remarkable performance in predicting pathological lesions of interest.

Background Liver fibrosis is an early stage of liver cirrhosis. As a reversible lesion before cirrhosis, liver failure, and liver cancer, it has been a target for drug discovery. Many antifibrotic candidates have shown promising results in experimental animal models; however, due to adverse clinical reactions, most antifibrotic agents are still preclinical. Therefore, rodent models have been used to examine the histopathological differences between the control and treatment groups to evaluate the efficacy of anti-fibrotic agents in non-clinical research. In addition, with improvements in digital image analysis incorporating artificial intelligence (AI), a few researchers have developed an automated quantification of fibrosis. However, the performance of multiple deep learning algorithms for the optimal quantification of hepatic fibrosis has not been evaluated. Here, we investigated three different localization algorithms, mask R-CNN, DeepLabV3+, and SSD, to detect hepatic fibrosis. Results 5750 images with 7503 annotations were trained using the three algorithms, and the model performance was evaluated in large-scale images and compared to the training images. The results showed that the precision values were comparable among the algorithms. However, there was a gap in the recall, leading to a difference in model accuracy. The mask R-CNN outperformed the recall value (0.93) and showed the closest prediction results to the annotation for detecting hepatic fibrosis among the algorithms. DeepLabV3+ also showed good performance; however, it had limitations in the misprediction of hepatic fibrosis as inflammatory cells and connective tissue. The trained SSD showed the lowest performance and was limited in predicting hepatic fibrosis compared to the other algorithms because of its low recall value (0.75). Conclusions We suggest it would be a more useful tool to apply segmentation algorithms in implementing AI algorithms to predict hepatic fibrosis in non-clinical studies.

Deep learning has recently become one of the most popular methods of image analysis. In non-clinical studies, several tissue slides are generated to investigate the toxicity of a test compound. These are converted into digital image data using a slide scanner, which is then studied by researchers to investigate abnormalities, and the deep learning method has been started to adopt in this study. However, comparative studies evaluating different deep learning algorithms for analyzing abnormal lesions are scarce. In this study, we applied three algorithms, SSD, Mask R-CNN, and DeepLabV3+, to detect hepatic necrosis in slide images and determine the best deep learning algorithm for analyzing abnormal lesions. We trained each algorithm on 5750 images and 5835 annotations of hepatic necrosis including validation and test, augmented with 500 image tiles of 448 × 448 pixels. Precision, recall, and accuracy were calculated for each algorithm based on the prediction results of 60 test images of 2688 × 2688 pixels. The two segmentation algorithms, DeepLabV3+ and Mask R-CNN, showed over 90% of accuracy (0.94 and 0.92, respectively), whereas SSD, an object detection algorithm, showed lower accuracy. The trained DeepLabV3+ outperformed all others in recall while also successfully separating hepatic necrosis from other features in the test images. It is important to localize and separate the abnormal lesion of interest from other features to investigate it on a slide level. Therefore, we suggest that segmentation algorithms are more appropriate than object detection algorithms for use in the pathological analysis of images in non-clinical studies.