Real-Time Abdominal Trauma Detection Using LSTM Neural Networks with MediaPipe and OpenCV Integration

Abstract

This study focused on developing an advanced system for abdominal trauma detection using an LSTM-based neural network. The objectives include achieving high accuracy in trauma detection, minimizing false positives and negatives, and providing real-time feedback to enhance patient outcomes. Current systems face challenges such as limited datasets, variability in real-world applications, and high computational demands, which hinder their effectiveness and generalizability. The methodology involved integrating MediaPipe for keypoint detection, OpenCV for real-time video processing, and a robust data preprocessing pipeline to train the LSTM model. The model demonstrated promising results, achieving 92% accuracy, 90% precision, 88% recall, and an F1 score of 89%, thus significantly outperforming the baseline models. In addition, the study explored multimodal fusion techniques to incorporate additional sensory inputs, further enhancing interpretative capabilities. The findings suggest that the proposed system can effectively detect abdominal trauma, offering a substantial improvement over the existing methods. Achievements include the successful deployment of a real-time detection system and development of a comprehensive dataset for training and evaluation.