Development of a deep learning model to automatically localize anatomic landmarks important for intraperitoneal free fluid detection during focused assessment with sonography in trauma (FAST)

Background The Focused Assessment with Sonography in Trauma (FAST) exam allows clinicians to rapidly identify intraperitoneal free fluid (FF), which can accumulate in several discrete anatomic locations. These locations represent potential spaces and are defined by adjacent key organ features. Satisfactory evaluation of these organ features is necessary to obtain a complete FAST exam. We sought to develop an algorithm that accurately identified relevant organs important for the FAST exam. Methods Data was used from 218 prospectively enrolled subjects with positive and negative FAST exams yielding 2,218 videos and 323,304 frames. Images were frame annotated for the presence of the following key organ features: liver, caudal liver tip, spleen, caudal spleen tip, kidney, diaphragm, and bladder. Data were randomly split into training, validation, and test sets in a 6:2:2 ratio for training a deep learning model using the YOLOv5 object detection algorithm. Diagnostic accuracy of the test set is reported at the frame and video levels. Results At the frame level, the mean average precision (mAP 50 ) was 0.71. At the video level, the mean F1 score for organ detection was 0.85. Video level detection of anatomic locations (defined by the co-incidence of two relevant organ structures) of potential FF accumulation showed a mean F1 of 0.79. At the video level, the model was optimized to favor a higher mean specificity (0.90, 95% CI 0.89–0.92) at the expense of a slightly reduced mean sensitivity (0.88, 95% CI 0.85–0.90). The inference speed was 26 ms per frame (38 frames per second) on a mobile tablet. Conclusions This organ detection algorithm identifies the key organ features and their interfaces, which are important for FF assessment. The algorithm has a low inference speed that approximates real time temporal resolution, providing a foundation for future work supporting automated image capture and quality assessment.

Authors

Institutions

Publication Details

Journal
PLoS ONE
Published
2026-11-19
DOI
https://doi.org/10.1371/journal.pone.0360536
Primary Topic
Ultrasound in Clinical Applications
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Development of a deep learning model to automatically localize anatomic landmarks important for intraperitoneal free fluid detection during focused assessment with sonography in trauma (FAST)

Goutam Ghoshal, Aaron M. Silver, Maxwell A. Thompson, Martha Grewe Wilson et al.
PLoS ONE
Ultrasound in Clinical Applications
article

Development of a deep learning model to automatically localize anatomic landmarks important for intraperitoneal free fluid detection during focused assessment with sonography in trauma (FAST)

Goutam Ghoshal, Aaron M. Silver, Maxwell A. Thompson, Martha Grewe Wilson, Cristiana Baloescu, David Oren Kessler, Mohsen Zahiri, Nikolai Schnittke, Alfredo Sabbaj, HyeonWoo Lee, Bryson Hicks, Julie A Rizzo, Cynthia R. Gregory, Aishwarya Sreenivasan, Balasundar I. Raju, Caelan Thomas, Stephen Schmidt, Kenton W. Gregory, Christopher Moore, Rashid Al Mukaddim, Guohao Chen, Michael Petrovich, Maggie Feuerherdt, Yuan Zhang, Maria Parker
article en

Abstract

Background The Focused Assessment with Sonography in Trauma (FAST) exam allows clinicians to rapidly identify intraperitoneal free fluid (FF), which can accumulate in several discrete anatomic locations. These locations represent potential spaces and are defined by adjacent key organ features. Satisfactory evaluation of these organ features is necessary to obtain a complete FAST exam. We sought to develop an algorithm that accurately identified relevant organs important for the FAST exam. Methods Data was used from 218 prospectively enrolled subjects with positive and negative FAST exams yielding 2,218 videos and 323,304 frames. Images were frame annotated for the presence of the following key organ features: liver, caudal liver tip, spleen, caudal spleen tip, kidney, diaphragm, and bladder. Data were randomly split into training, validation, and test sets in a 6:2:2 ratio for training a deep learning model using the YOLOv5 object detection algorithm. Diagnostic accuracy of the test set is reported at the frame and video levels. Results At the frame level, the mean average precision (mAP 50 ) was 0.71. At the video level, the mean F1 score for organ detection was 0.85. Video level detection of anatomic locations (defined by the co-incidence of two relevant organ structures) of potential FF accumulation showed a mean F1 of 0.79. At the video level, the model was optimized to favor a higher mean specificity (0.90, 95% CI 0.89–0.92) at the expense of a slightly reduced mean sensitivity (0.88, 95% CI 0.85–0.90). The inference speed was 26 ms per frame (38 frames per second) on a mobile tablet. Conclusions This organ detection algorithm identifies the key organ features and their interfaces, which are important for FF assessment. The algorithm has a low inference speed that approximates real time temporal resolution, providing a foundation for future work supporting automated image capture and quality assessment.

PLoS ONEVol. 21(11)
Brooke Army Medical Center (US), Oregon Health & Science University (US), University of Alabama at Birmingham (US), Yale University (US), Columbia University (US)
Ultrasound in Clinical Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon