Related Experiment Videos
Deep learning-based differentiation of FLASH ultrasound windows
Laura Marlene Köninger1, Tobias Müller2, Stefan Gesell-May2
1Equine Hospital, Ludwig-Maximilians-University Munich, Munich, Germany.
Background:
Equine colic emergencies present highly stressful, time-critical scenarios for both clinicians and patients, thus rapid and accurate diagnostic triage is paramount. To support veterinarians under these high-pressure conditions, deep learning tools can be leveraged for automated image recognition to assist in localizing abdominal organs and interpreting FLASH (Fast Localized Abdominal Sonography of Horses) ultrasound windows.
Objectives:
A software tool was developed to identify FLASH-specific ultrasound windows on the equine abdomen and to recognize ultrasonographic organ patterns.
Study Design:
Retrospective-prospective comparison of software and clinical examination, model development.
Methods:
A deep learning approach for image classification was used to train software to assign ultrasound images to FLASH windows.
Results:
Strongest software performance was observed for the duodenal window, with all test images correctly classified, corresponding to a recall of 1.00. Nephrosplenic window demonstrated high classification accuracy, but 23% of images were misclassified as gastric window. While 84% of ventral images were correctly classified, 16% were predicted as duodenal window. Most pronounced confusion was observed between ventral and thoracic windows, with only 55% of thoracic images correctly identified. Left and right middle third FLASH windows 4 and 6 were substantially underrepresented and were therefore excluded from further analyses. In addition, these windows exhibited high anatomical similarity, which further complicated their differentiation.
Main Limitations:
Sources of selection bias was interpretation of windows and the definition of normal physiological ultrasound images by the examiners. The specific FLASH windows were not homogeneously dispersed in the dataset (e.g., FLASH windows 4 and 6 were underrepresented).
Conclusions:
The developed deep learning tool demonstrates the feasibility of automated recognition of standardized equine FLASH windows and may support ultrasound training as an educational aid for first-day competencies. Future incorporation of organ-level annotations and probe position information may improve discrimination between anatomically similar windows and enhance model performance.