Philippine E-Journals

INVESTIGATION ON THE EFFECTS OF RADIOGRAPHIC IMAGE QUALITY ATTRIBUTES ON THE PERFORMANCE OF CONVOLUTIONAL NEURAL NETWORKS (CNNS) IN DETECTING COVID-19

Reymond R Mesuga | Cloyd Raymond P Pernes | Luther A Villacruz | Mark Anthony C Burgonio

Discipline: bioengineering, medical and biomedical engineering

 

Abstract:

Radiographic image quality is one of the factors that impacts professionals’ decisions when diagnosing lung diseases using X-ray images. Hence, poor radiographic image quality could result in a misleading diagnosis affecting the person being investigated. This is true in human vision, as well as the computer vision. This study investigated the effects of different radiographic image quality attributes (i.e., contrast, Gaussian blur, Gaussian noise, and salt-and-pepper noise) on the performance of various Convolutional Neural Networks (CNNs) models. We use COVID-19 x-ray data as an initiative to the pandemic, apply different radiographic image quality attributes, and test the performance of CNN models in the effects of the attributes in the classification task. The results showed the following: (i) increasing levels of experimented noises (i.e., Gaussian and salt-and-pepper noise) rapidly decreases the performance of the models with no sign of resiliency; (ii) decreasing contrast appears to be beneficial at some particular level (e.g., contrast factor = 3); and (iii) increasing Gaussian blur decreases the performance of models but less rapidly than that of noises. As a conclusion, increasing noise like Gaussian and salt-and-pepper noise can be considered as a hindrance to the performance of CNNs while decreasing contrast and increasing Gaussian blur seemed to be beneficial especially if applied for data augmentation or enhancement techniques as the performance of the CNNs were observed to be more resilient against these two attributes than that of noises.