Inspired by “Focus, Fusion, Collaboration”: A multi-level ensemble network for automatic pneumonia diagnosis from full slice CT images

Abstract

Pneumonia is an infectious disease that endangers human health. With advancements in science and technology, deep learning-driven techniques have gained prominence in this field. However, their applicability to clinical practice remains limited because they mostly neglect three key points: focus on local lesion regions, multi-level feature fusion, and sequential collaborative decision-making. In this paper, we present a novel multi-level ensemble network for automatic pneumonia diagnosis from full slice CT images, inspired by the “Focus, Fusion, Collaboration” strategy. Our proposed model involves three modules: the global–local feature extraction module is first designed to fully extract the global structure information and local lesion details; subsequently, the multi-level feature fusion module is responsible for integrating the above-mentioned global and local information; finally, the sequential pneumonia prediction module is utilized to learn the contextual relationship between the adjacent slices, thus generating the final diagnosis results. Building upon mimicking the diagnostic behavior from real-world clinical scenarios, our model enables the integration of multiple types of information (including global structure information, local lesion features, and slice dependencies) and sequential pneumonia diagnosis. Extensive comparative experiments are conducted to verify the feasibility and effectiveness of our proposed method. The experimental results show that our model can obtain an accuracy of 91.4% in a four-class pneumonia diagnosis task, outperforming the other classical works.