The value of artificial intelligence (AI) in detection of post traumatic brain injury using non-contrast CT scans

Background: Traumatic brain injury is a leading cause of morbidity and mortality worldwide. Rapid and accurate assessment of TBI is crucial for timely intervention. Non-contrast computed tomography is the primary imaging modality for initial assessment. Aim of this study: To assess the role of artificial intelligence in the detection of post traumatic brain injury using non-contrast CT scans using a deep learning model from U-Net and resnet50 architectures. Methods: A cross-sectional study was conducted using 628 patients, divided into training, testing and validation sets. A deep learning model, utilizing U-Net for semantic segmentation and ResNet-50 for feature extraction, was trained and validated for detection of PTBI, and was compared to two radiologists on the validation dataset of 1763 images. Results: The AI model demonstrated a sensitivity of 94.3% and a specificity of 84.6% in detecting PTBI when compared to the first radiologist's findings. Agreement between the first radiologist and AI in final diagnosis yielded a Kappa value of 0.796 (p < 0.001). The AI model achieved an AUC of 0.976 (p < 0.001) for detecting abnormalities compared with the first radiologist. Similarly, the comparison between the second radiologist and AI showed a sensitivity of 94.5%, specificity of 83.5%, and a kappa value of 0.787 (p < 0.001), with AUC of 0.968 (p < 0.001). Conclusion: This study demonstrates the potential of deep learning models to accurately detect PTBI on NCCT scans, with diagnostic performance comparable to expert radiologists. AI can aid radiologists in prioritizing critical cases and reducing diagnostic time.