Evaluation of Depth-wise Separable Convolution and Channel Attention Mechanism to Bacilli Segmentation

Abstract

Tuberculosis (TB) is the second leading cause of death from a single infectious agent. In the field of computer vision, much works has been developed to automate the identification of bacilli in sputum smear microscopy field images. In recent years, deep machine learning techniques have been increasingly used. In this work, architectures with different depth, implemented with channel attention mechanism and with depth-wise separable convolution are evaluated to accomplish the task of bacilli segmentation in microscopy images. The best result was obtained with a deeper architecture, implemented with channel attention mechanism and depth-wise separable convolution, with an F1-score of 99.38%.