Purpose or Case Report: CXR is the most common imaging method to diagnose pneumonia in children in limited-resource settings. There is a need to simplify and expedite its interpretation. By using a machine learning model to first classify and interpret the pneumonia images and then incorporate those characteristic imaging findings patterns into a simulated mobile app, health care workers can use their mobile devices to interpret those findings based on preloaded images built into their mobile devices corresponding to pneumonia.
Methods & Materials: We used 6543 CXR images from the PERCH (Pneumonia Etiology Research for Child Health) project containing findings compatible with pneumonia as a background to train a model to identify abnormal findings in children 0-59 months. The model was created using a ResNet-50 convolutional neural network to classify the images. A modified version of the network was trained using images from the database. This model was optimized for mobile deployment using the Tensor Flow Lite Task Library and integrated on a simulated mobile application, which allows for pneumonia findings detected from a CXR image to be identified through the phone’s camera. The model containing the already created patterns to recognize pneumonia can be incorporated into a real mobile device application.
Results: The model achieved 91.12% accuracy, 90.7% specificity and 96.0% sensitivity to detect pneumonia findings present in the CXRs. This performance was further tested by placing CXRs in the android device simulated environment to test probable real world camera usage behaviors.
Conclusions: Accurate interpretation of CXR findings compatible with pneumonia in children is a challenge in limited-resource settings. Expediting the process of such interpretation to diagnose pneumonia efficiently and on time can be done with a sensitivity of 96.0% by incorporating previously loaded artificial intelligence recognition patterns from previously done CXRs into a mobile phone app.