ETDHDNet: An advanced DenseNet-based extended texture descriptor for efficient tuberculosis prediction in CXR images

Chest X-ray imaging (CXR) for the prediction of tuberculosis (TB) is essential in medical diagnostics, as it plays a vital role in early detection and the development of effective treatment plans. Although Convolutional Neural Networks (CNNs) are effective in extracting features from images in a hierarchical manner, they have limitation of texture-related features due to their emphasis on capturing global patterns. To address this shortcoming, we propose an Extended Texture Descriptor Histogram DenseNet (ETDHDNet), a model designed for TB prediction from CXR images by integrating texture analysis with deep feature learning. ETDHDNet consists of three key components: the Extended Texture Descriptor Histogram (ETDH) module to capture multi-scale texture features across fine, medium, and coarse granularities; a hierarchical feature learning unit with densely connected layers for extracting high-level features; and a neural network for handling binary and multi-class TB prediction. Experimental results demonstrate that ETDHDNet surpasses existing methods in TB prediction from CXR images across three datasets. For binary classification, the model achieves an AUC of 0.998 on TBX11K, 0.997 on the Tuberculosis Chest X-ray Database, and 0.930 on the Shenzhen Dataset, as well as an AUC of 0.993 for multi-class TB prediction on TBX11K. • Proposed ETDHDNet, a model for tuberculosis prediction from CXR images. • Introduced ETDH module to capture fine, medium, and coarse texture details in CXR images. • Developed a dual feature extraction method combining ETDH with a tailored DenseNet121. • Evaluated ETDHDNet on binary and multi-class tasks across three TB CXR datasets, improving prediction performance.