A novel stacked generalization of models for improved TB detection in chest radiographs

Abstract

Chest x-ray (CXR) analysis is a common part of the protocol for confirming active pulmonary Tuberculosis (TB). However, many TB endemic regions are severely resource constrained in radiological services impairing timely detection and treatment. Computer-aided diagnosis (CADx) tools can supplement decision-making while simultaneously addressing the gap in expert radiological interpretation during mobile field screening. These tools use hand-engineered and/or convolutional neural networks (CNN) computed image features. CNN, a class of deep learning (DL) models, has gained research prominence in visual recognition. It has been shown that Ensemble learning has an inherent advantage of constructing non-linear decision making functions and improve visual recognition. We create a stacking of classifiers with hand-engineered and CNN features toward improving TB detection in CXRs. The results obtained are highly promising and superior to the state-of-the-art.

Figure 1.

CXR images showing 2 examples of pulmonary abnormalities (left: pleural effusion, middle: cavitary lung lesion right lung), and normal lung image (right).

Figure 2.

PA CXR lung ROI segmentation process: (a) original image, (b) computed lung mask, (c) segmented lung ROI with bounding box.

Figure 3.

Stacked generalization of models from the proposals.