DEEP LEARNING-BASED ALGORITHMS CAN BE USED TO PREDICT CYTOLOGIC TOTAL HEMOSIDERIN SCORES

Background: Exercise-Induced Pulmonary Hemorrhage (EIPH) is a common condition in horses that is associated with reduced race performance. The gold standard for diagnosis is cytologic quantification of hemosiderophages in bronchoalveolar lavage fluid using the Total Hemosiderin Score (THS). For this scoring system, 300 alvelolar macrophages are graded into 5 classes based on their hemosiderin content. However, routine application of the THS is hampered by the high time investment. Objective: To evaluate the performance of a deep learning-based algorithm for automated THS analysis. Methods: Fifty-two cytospin preparations stained with special iron stains were collected from 26 horses with clinically suspected EIPH. The THS was determined in whole-slide images by ten pathologists (scoring 300 cells) using annotation software (SlideRunner or EXACT) and a deep learning-based algorithm (scoring all cells in the whole-slide images). Results: Correlation of the mean pathologists’ THS (r = 0.98) or the individual pathologist’s THS (r ranging between 0.95 and 0.98) with the algorithmic THS had almost perfect agreement. Accuracy for determining a value above the diagnostic THS cut-off of 75 (as compared to a ground truth dataset) was 92.3% for the algorithm and 63.4-92.3% for the individual pathologists. Whereas the algorithm analyzed entire whole-slide images in less than 2 minutes, pathologists required on average 15 minutes to score 300 hemosiderophages. Conclusion: As compared to the pathologists, the evaluated algorithms had very high accuracy and reproducibility. Deep learning-based algorithms are a promising tool to facilitate widespread use of the THS for routine diagnostics.