This year we updated our CXR screening algorithms to include Convolutional Neural Network-based (CNN) Deep Learning models that resulted in improvements in classification performance. We studied several pre-trained off-the-shelf deep learning frameworks and compared their performance against a custom designed CNN architecture. The outcomes were visualized with a novel visualization algorithm; and then field-tested on a CXR screening system deployed in rural parts of western Kenya through collaboration with Indiana University, AMPATH (a Kenyan NGO); and an expert radiologist from the University of California, San Francisco. We expanded our interest in CXR disease classification to include pneumonia in both adults and children, and also evaluated ability of artificial intelligence algorithms in detecting drug-resistant variants of TB in chest x-rays. There is a need for further work in separating pulmonary opacity pneumonia from TB, particularly in HIV-positive children. Further, we showed that it is possible to separate drug resistant from non-drug resistant TB in adults using artificial intelligence, albeit with a lower confidence. We also studied the ability of deep learning algorithm in detecting and staging the severity of cardiomegaly using CXRs. We also studied the use of generative adversarial networks (GANs) in a pilot study to evaluate their ability to produce deepfake CXR images to increase the number of training images with a goal to improve deep learning-algorithm based classification outcomes. We found the results to be mediocre toward meeting this goal. Further study is needed on the statistical variation needed in image datasets to improve artificial intelligence algorithm outcomes. Finally, we studied the use of adversarial networks in automatically generating radiology reports.
| Karargyris, Alexandros; Siegelman, Jenifer; Tzortzis, Dimitris et al. (2016) Combination of texture and shape features to detect pulmonary abnormalities in digital chest X-rays. Int J Comput Assist Radiol Surg 11:99-106 |
