Improving pneumonia diagnosis accuracy via systematic convolutional neural network-based image enhancement

Chest X-rays play a significant role in diagnosing pneumonia due to the technology's cost-effectiveness and rapid development times. Detecting pneumonia in chest Xrays is a challenging process that relies heavily upon the availability of trained radiologists and high-quality imagery. Training qualified interpreters require significant resources, while medical imaging remains prone to a wide variety of deficiencies. Therefore, an automated system for pneumonia diagnosis consisting of three phases is proposed. An initial sorting phase consisting of a trained ResNet-18 convolutional neural network separates the dataset according to the interpretive quality of the images, creating a high and low-quality class. The unique image translation capabilities of the CycleGAN network are leveraged in the enhancement phase to translate low-quality images into improved versions. A final ResNet-18 network serves to classify pneumonia in the diagnosis phase. The enhancement system improved mixed quality diagnosis accuracy by 12.1% to 86.7%, with training sets composed of enhanced images achieving an accuracy 15.8% higher than their low-quality counterparts. The system's generalized method for image augmentation successfully mitigates the deficiencies of low-quality data, allowing for a higher accuracy diagnosis than otherwise possible.