T-Fusion Net: A Novel Deep Neural Network Augmented with Multiple Localizations Based Spatial Attention Mechanisms for Covid-19 Detection

Abstract

Recent years have witnessed significant advancements in image classification tasks, primarily driven by the increasing capabilities of deep neural networks. Nonetheless, the growing complexity of datasets and the ongoing pursuit of enhanced performance necessitate innovative approaches. In this study, we introduce a novel deep neural network, referred to as the “T-Fusion Net,” which incorporates multiple spatial attention mechanisms based on localizations. This attention mechanism enables the model to concentrate on pertinent regions within the images, thus bolstering its discriminative abilities. To further elevate image classification accuracy, we employ a homogeneous ensemble of these T-Fusion Nets. This ensemble technique involves multiple instances of individual T-Fusion Nets, and the fusion of their outputs is achieved through a fuzzy max fusion process. We meticulously optimize this fusion process by selecting appropriate parameters to ensure a balanced contribution from each individual model. Experimental assessments conducted on a well-documented dataset of COVID-19 (SARS-CoV-2 CT scans) were utilized to assess the efficacy of both the T-Fusion Net model and its ensemble counterpart. The results indicate that our T-Fusion Net and its ensemble model consistently surpass existing approaches, demonstrating remarkable accuracy rates of 97.59% and 98.4%, respectively.