Developing Deep Learning Models to Diagnose Lung Opacities in Photos of Chest X-rays

Abstract

Machine learning models have the potential to streamline the diagnosis of urgent health conditions related to the chest region, ranging from an enlarged heart to pneumonia. In order to integrate machine learning models into clinical practice where radiologists are not available, recent approaches have aimed to develop models on photos of chest X-rays so that clinicians can quickly receive machine learning model-based diagnoses through an app on their mobile device. In my work, I developed convolutional neural networks on photos of chest X-rays in order to automatically diagnose lung opacities. The models were trained using the CheXphoto dataset, a large-scale collection of many smartphone photos and synthetic photographic transformations of chest X-rays. The use of ImageNet pre-training improved model AUC by a margin of around 5 percent, and the use of data augmentation did not improve the performance of the model on the validation set. The Xception architecture with ImageNet pre-training and no data augmentation achieved the highest performance on the validation set and attained an AUC of 0.8580 on the  test set.