208 - Artificial Intelligence to Detect Spinal Fluid and Spinal Cord on Ultrasound

Poster Number: 208
Publication Number: 208.2124

Background: Lumbar puncture (LP) is a critical diagnostic test in febrile infants that is associated with high failure rates, especially among novice providers. Ultrasound (US) guidance can improve LP success but is underutilized due to lack of provider comfort and expertise in this skill. Automated identification of optimal interspaces on US has the potential to improve procedural success.

Objective: Our aim was to develop an artificial intelligence (AI) algorithm using a database of ultrasound spinal anatomy videos to identify key anatomic structures and aid in infant LP performance.

Design/Methods: 5 patient videos were included contributing 1515 total frames to model development and testing. Frames were annotated with binary classification for presence of key anatomical features including spinal cord and spinal fluid and determining whether the image is bad or good quality. We treated the problem as a multi-label classification task at the frame level. Data were augmented to increase dataset to 11224 frames and to balance classes between training and testing sets. Image processing techniques were used to enhance image features for model learning. We employed deep learning architectures including ResNet18, ResNet34, AlexNet, and DenseNet, with hyperparameter tuning for optimal performance. We evaluated test characteristics of all models to identify individual features. F1 score was computed to assess for optimal balance between true positives, false positives, and false negatives.

Results: The ResNet18 model performed best with an average f1 score of 0.98 for all three feature classifications. Table 1 shows the test characteristics for model identification of spinal cord and spinal fluid. Figure 1. Shows examples of correct and incorrect classifications.

Conclusion(s): The model successfully identified images with spinal fluid, spinal cord, and of poor quality with high accuracy using ResNet architecture. Future work will expand the dataset, identify additional features, and label pixels for specific features.