An efficient deep learning-based morphology aware hierarchical mixture of features for tuberculosis screening using segmentation of chest X-ray images

Tuberculosis (TB) is a chronic lung disorder caused by bacterial infection and is a major cause of death. Lung cancer also has a significant impact, and existing solutions concentrate on initial screening, which mainly results in better outcomes at a comparatively lower cost. Screening, particularly by chest X-rays (CXR), is globally recognized as an effective method for reducing lung cancer mortality. Therefore, a precise and initial identification of TB is highly crucial, or else, it threatens lives. In the investigation into cases of TB, CXR images are not only the primary method of diagnosis according to medical imaging, but also the radiological diagnosis. The recent developments of computing, deep learning (DL), for image processing, carry a beneficial effect for the automated identification of numerous illnesses from CXRs. Now, the effectiveness of lung segmentation and TB screening methods is established for CXRs analysis by the DL technique to help radiologists recognize suspicious lesions and nodes in lung cancer patients. This paper presents an Efficient Deep Learning-Based Hierarchical Feature Fusion Approach for Lung Segmentation and Tuberculosis Screening (EDLHFFA-LSTS) model. The aim is to develop an automatic DL-based framework for precise lung segmentation and TB screening using CXR images to support early diagnosis and clinical decision-making. Initially, the image pre-processing stage includes resizing, adaptive filtering (AF), and histogram equalization (HE) to enhance the image quality. For the segmentation process, the EDLHFFA-LSTS model implements the Res-UNet method. Furthermore, the fusion of EfficientNetV2, CapsNet, and Convolutional Vision Transformer (CViT) techniques is employed for the feature extraction process. Finally, the stacked autoencoder (SAE) technique is implemented for classification. Extensive simulations were conducted to demonstrate the promising results of the EDLHFFA-LSTS methodology on the CXR Masks and Labels dataset. The comparison study of the EDLHFFA-LSTS methodology illustrated a superior accuracy value of 98.33% over existing models.