Publikationsdatenbank

Scoring exercise-induced pulmonary hemorrhage:
accuracy of human observer and deep-learning based algorithms (2019)

Art

Vortrag

Autoren

Bertram, Christof A. (WE 12)

Marzahl, Christian

Aubreville, Marc

Stayt, Jason

Jasensky, Anne-Katherine

Bartenschlager, Florian (WE 12)

Fragoso-Garcia, Marco (WE 12)

Jabari, Sami

Elsemann, Svenja

Barton, Ann K. (WE 17)

Maier, Andreas

Hill, Jenny

Klopfleisch, Robert (WE 12)

Kongress

ACVP & ASVCP 2019 Concurrent Annual Meeting

San Antonio, 09. – 13.11.2019

Quelle

Sprache

Englisch

Verweise

URL (Volltext): https://cdn.ymaws.com/www.acvp.org/resource/resmgr/meetings_&_events/2019/2019_acvp_and_asvcp_final_ab.pdf

Kontakt

Pferdeklinik

Oertzenweg 19 b
14163 Berlin
+49 30 838 62299 / 62300
pferdeklinik@vetmed.fu-berlin.de

Abstract / Zusammenfassung

Tuesday, November 12, 20198:30 a.m. –8:45 a.m.

Background:
Exercise-induced pulmonary hemorrhage (EIPH)is a common disease of racehorses with strenuousexercise. There is good evidence that this disease has negative impact on athletic performance. The gold standard diagnostic method is cytological examination and scoring of pulmonary hemosiderophages (PH). An established scoring system requires grading of 300 PH into five grades depending on the degree of cytoplasmic hemosiderin storage.

Objective:
To determine accuracy of cytological scoring of PH by human observers and deep learning-based algorithms.

Methods:
A ground truth dataset of 17 completely annotated whole slide images (WSI) of equine bronchoalveolar lavage stained for iron content was developed. Accuracy of eight human observers and the classification algorithm were determined in 2000 randomly-selected PH. Precision of the object detection algorithm was tested in three WSI.

Results:
Classification of participants had an intra-observer concordance of 68-88% with Cohen’s κ ranging from 0.60–0.84. Mean consistency with the ground truth dataset was 0.73%. The algorithm achieved a κ=1.0 and classified85% of the 2000 cells in agreement with the dataset.
The object detection algorithm had a mean average precision of 0.66 with an average error in the grade of 0.09. Calculation time of entire whole slide images was <2 minutes.

Conclusion:
Manual scoring is not only monotonous and time-consuming, but also exhibits high degree of inter-and intra-observer variability. On the other hand, the algorithmic approach has an accuracy comparable to human experts, 100% reproducibility and short calculation times. Therefore, we consider image analysis as a feasible solution for quantification of EIPH.