See through the noise: revolutionizing medical image diagnosis with quadratic convolutional neural network (Q-CNN)

Abstract

This study introduces a novel approach that utilizes quadratic convolutional neural networks (Q-CNN) to enhance the sensitivity of neural network models in analyzing noisy radiographs without the need for training on noisy images. The Q-CNN model is applied for COVID-19 classification in medical image diagnosis. In the presence of noise, the Q-CNN model exhibits superior performance compared to several benchmark models for COVID-19 diagnosis. The Q-CNN model effectively detects and classifies specific indicators, showcasing a unique advantage over existing models. Experimental evaluations conducted on chest radiographs with varying levels of noise demonstrate that the Q-CNN surpasses benchmark models in effectively handling noisy images while maintaining high classification accuracy. The visualization of features and the generation of heatmaps shed light on the essential role played by the non-linear expansion and cross-correlation mechanisms within the Q-CNN in overcoming limitations imposed by noise. Furthermore, a similarity analysis confirms the noise-resistance capabilities of the Q-CNN, providing further validation of its effectiveness in identifying and classifying indicators within noisy radiographs. This research significantly contributes to the field of medical diagnosis by presenting a reliable tool for the detection of specific indicators in noisy radiographic images, thereby enhancing accuracy and improving patient care.