Ensemble Learning With Attention-Based Multiple Instance Pooling for Classification of SPT

Abstract

Tuberculosis is one of the most common infections in the human population, while reactivation of the MTB bacteria can lead to secondary pulmonary tuberculosis (SPT). SPT is a significant health risk for both immunocompromised individuals and the population of regional hotspots. This paper proposes a novel model for ensemble learning for the efficient identification of SPT in CT imaging. We reformulate the ensemble process as multiple instance learning and utilizes an attention-based pooling layer to ensemble features extracted from a range of base networks. The best performing model consists of 4 base models and an ensemble network with Hopfield-pooling. We abbreviate this model as Hopfield-pooling ensemble (HFPE). For the classification of CT slices between healthy controls and SPT patients, HFPE achieved an accuracy of 96.00±2.55%, specificity of 97.00±2.45%, precision of 96.99±2.46%, sensitivity of 95.00±4.47, and F1-score of 95.92±2.63%, achieving comparable or higher performances to a number of state-of-art approaches.