R. A. Callcut1,2, M. Girard2, S. Hammond2, T. Vu2,3, R. Shah2,3, V. Pedoia2,3, S. Majumdar2,3 1University Of California – San Francisco,Surgery,San Francisco, CA, USA 2University Of California – San Francisco,Center For Digital Health Innovation,San Francisco, CA, USA 3University Of California – San Francisco,Radiology,San Francisco, CA, USA
Introduction: Chest Xrays (CXR) are the most common imaging modality used worldwide. The time from image ascertainment until review creates an inherent delay in identification of potentially life threatening findings. Bedside deployed or point of care (POC) tools developed from neural networks have potential to speed the time to clinician recognition of critical findings. This study applies neural networks to CXRs to automate the detection of peumoperitoneum (PP).
Methods: We utilized a multi-step deep learning pipeline to create a clinical decision support system for the detection of pneumoperitoneum under the right diaphragm. 5528 training and 1368 validation images were used to train a Unet to initially segment the right and left lung. By combining the lung segmentation with simple, rule-based algorithms we generated a region of interest in the original image where important features of positive-case detection were likely to be found (Figure 1a). Two readers blindly read images in a second clinical dataset (1821 CXRS total with 771 positive for PP) to classify PP presence or absence. Images were then divided randomly into a 75% training, 15% validation, and a 10% testing sets. With the cropped, full resolution images of the region of interest (Figure 1b), a DenseNet neural network classifier was trained to identify PP.
Results: The AUROC for training was 0.99, validation 0.95, and testing 0.96 (Figure 1c). This yielded a specificity of 94% in the validation group and the results remained consistent in the testing set (92%). Overall, the accuracy for detection of PP exceeded 90% in the validation group and was confirmed to be excellent in the testing set (92%).
Conclusion: This work demonstrates the potential power of integration of Artificial Intelligence into POC clinical decision support tools. These algorithms are highly accurate and specific and could yield earlier clinician recognition of potential life threatening findings.
