Postdoctoral fellow Dr. Yuxing Tang and colleagues created the artificial intelligence (AI) approach using three main modules -- a U-Net autoencoder, a convolutional neural network (CNN) discriminator, and an encoder -- that simultaneously learn and collaborate with each other to carry out the task.
The AI training process began with the trained model seeing only normal chest x-rays. In this manner, the model is intentionally trained to reconstruct normal chest x-rays but perform poorly on abnormal chest x-ray reconstruction.
Tang and colleagues found that their AI model could reconstruct normal chest x-rays of good quality in the testing stage. Conversely, the network was unable to reconstruct abnormal chest x-rays of good quality, which was the desired outcome.
Using this AI learning strategy, the algorithm might reduce the need for manual annotations in the clinical setting, the researchers concluded, "thereby accelerating the development of automated chest x-ray diagnosis software for radiologists." They also suggested that this approach might be extended to other medical image modalities.
