High-accuracy eosinophil detection in eosinophilic esophagitis histological images using machine learning model YOLO11

Objective. To evaluate the effectiveness of the YOLO11 machine learning model for automated segmentation and detection of eosinophils in histological images in order to improve the diagnostic accuracy of eosinophilic esophagitis (EoE). Materials and Methods. A multicenter retrospective analysis was conducted using histological images obtained through whole slide imaging (WSI) from 60 patients diagnosed with EoE. Out of 653 tissue section images, 54 were manually annotated and reviewed. The annotated dataset was then used to train the YOLO11 model. Results. By the 150th training epoch, the model demonstrated consistent improvement in precision and recall for bounding boxes detection. The final recall value reached 0.98, indicating a very high sensitivity of the model to the target areas (i.e., eosinophils) within the histological slides. The Intersection over Union (IoU) score peaked at 0.94, reflecting the model’s robust performance in accurately localizing and segmenting eosinophils. Furthermore, the model exhibited strong precision and recall not only for bounding boxes but also for segmentation masks, reinforcing its capability to delineate eosinophil boundaries. However, performance evaluation on new, unannotated images revealed several limitations requiring further optimization, particularly a decrease in effectiveness when analyzing clusters of eosinophils. Conclusion. The YOLO11-based approach developed in this study represents a significant advancement in the automation of histological assessment for EoE, offering a highly accurate tool for eosinophilic infiltration analysis. Future research directions include expanding the annotated WSI dataset, refining the model through additional training, and exploring the application of vision transformer-based architectures. The results of eosinophil segmentation and peak eosinophil count (PEC) will require retrospective validation against values determined by pathologists through light microscopy.