Eosinophil Detection with Modified YOLOv3 Model in Large Pathology Image

Eosinophil detection and counting have an important role in the treatment of eosinophilic gastroenteritis. Generally, the physicians diagnose the eosinophilic gastroenteritis depending on the number of eosinophil cells in the corresponding pathology images. However, manually identifying eosinophils in the large pathology image is time-consuming and boring. The accuracy of Eosinophil counting significantly relies on the experiences of professional pathologists, the manual counting also prone to missing and misclassification due to fatigue. In this paper, we introduce a modified YOLOv3 object detection method to automatically count the eosinophil. Specifically, to detect the small and crowded eosinophils in the large pathology image, we carefully redesign the scale and size of candidate boxes and make our kernels more sensitive to tiny eosinophils in the feature maps. To verify our method, we collect and label a real eosinophil dataset with totally 8469 eosinophils being annotated and distributing over 255 pathology images from 11 different classes. We implemented our proposed method on our dataset and demonstrated the feasibility and bright prospect for applying object detection technique into the treatments of eosinophilic gastroenteritis.