Artificial Intelligence (AI) Solution for Plasma Cells Detection

Abstract

The article investigates the application of a neural network diagnosis model to histological images in order to detect plasma cells for chronic endometritis detection. A two-stage algorithm was developed for plasma cell detection. At the first stage, a CenterNet model was used to detect stromal and epithelial cells. The neural network was trained on an open dataset with histological images and further fine-tuned using an additional labeled dataset. A labeling protocol was used, and the coefficient of agreement between two experts was calculated, which turned out to be 0.81. At the second stage, using the developed algorithm based on computer vision methods, plasma cells were identified and their HSV color boundaries were calculated. For the two-stage algorithm the following quality metrics were obtained: precision = 0.70, recall = 0.43, f1-score = 0.53. The model then was modified to detect only plasma cells and trained on a dataset with histological images containing labeled plasma cells. The quality metrics of the modified detection model were obtained: precision = 0.73, recall = 0.89, f1-score = 0.8. As a result of the comparison, the modified detection model approach showed the best quality metrics. Automating the work of counting plasma cells will allow doctors to spend less time on routine activities.