Why Did AI Perform Better at Some Abdominal FAST Scan Sites Than Others?

Abdominal FAST examines several sites for free fluid, an approach established in early canine trauma research.[1] For clinicians considering artificial intelligence (AI), the practical question is whether an algorithm can interpret every window equally well. A March 2026 study published in Frontiers in Veterinary Science found substantial differences between models for different abdominal sites. Its results suggest that performance depends partly on the anatomy shown and the images used to train the model, although the study did not establish a single cause for those differences.[2]

Study objective

Martinez and colleagues collected ultrasound recordings from five healthy military working dogs and twenty client-owned dogs with effusions, pneumothorax or combinations of these findings. An experienced specialist reviewed and selected recordings obtained with one ultrasound system. The team developed separate models for the diaphragmatic-hepatic (DH) window, pooled splenorenal and hepatorenal views, and the cystocolic (CC) window.[2]

Key findings

The best DH model achieved mean sensitivity of 98% and accuracy of 97%. The combined renal model reached 83% sensitivity and 78% accuracy, while the CC model reached 84% and 77%, respectively. Sensitivity reflects the detection of positive images. Accuracy includes correct positive and negative classifications. These figures describe image classification in dogs held out from training, not the percentage of complete clinical examinations correctly diagnosed.[2]

The initial models failed to reach the study’s performance targets. A model pretrained on a more targeted ultrasound dataset containing bladder, uterine and ovarian images subsequently met those targets after further training and optimization. This supports the value of anatomically relevant training data, without proving that pretraining alone accounted for the improvement.[2]

Clinical utility

The authors suggested that similarities between liver-related pretraining images and the DH examination helped explain its stronger performance. The models used transfer learning, starting with features learned from existing human medical images before further training on canine ultrasound. Relevant prior learning may therefore have made fluid detection easier with relatively few canine examples.[2]

Fluid was not visible at every site in every affected dog. The renal dataset was also curated to retain the views that best represented each dog’s fluid status. The authors identified room for the model to learn more about the size and typical location of fluid near the kidneys. This suggests that recognizing fluid depends on its relationship to surrounding structures, and that the examples supplied during training matter.[2]

At the CC window, the model needed to distinguish fluid contained normally within the bladder from free peritoneal fluid outside it. The authors described variation in bladder size, positioning and surrounding fluid appearance as additional sources of complexity.[2]

Study limitations

The explanation for the DH result remains a proposed one. The best models differed in architecture and training, and the datasets differed between sites. The DH result cannot establish that this window is inherently superior for detecting fluid in every dog, or that anatomy alone produced the performance gap.

The researchers grouped the splenorenal and hepatorenal views because of their similar anatomy and appearance of effusion. Consequently, the reported figures cannot tell us whether the splenorenal window outperformed an individual hepatorenal window.[2]

False positives remained a limitation. Although CC sensitivity reached 84%, its area under the receiver operating characteristic curve was 0.66, compared with 0.995 for DH. This measure describes discrimination between positive and negative images across decision thresholds. A relatively high sensitivity therefore did not mean that the CC model reliably avoided false alarms.[2]

Image selection also matters. Low-quality recordings were excluded, and the DH dataset contained only eleven dogs. Performance varied across validation tests, so these averages should not be treated as expected results for every patient or image.[2]

Practice takeaways

The findings support assessing AI performance at each intended scan site. They do not justify reducing a FAST examination to the DH window. A negative image from one location cannot establish that fluid is absent elsewhere, particularly when the study itself found that fluid visibility varied between sites.[2]

For clinicians, the useful lesson is to ask which window a model was trained to interpret, which errors it makes there, and whether the image being assessed resembles its training examples. These results point toward the need for AI support adapted to each scan site, while preserving the complementary information gained from examining multiple abdominal windows.