Scholarlink

<p>Nontuberculous mycobacterial pulmonary disease (NTM-PD) is a progressive infectious illness with various radiological appearances such as bronchiectasis, tree-in-bud opacity, nodules, and parenchymal distortion. Manual diagnosis with 2D HRCT has been associated with substantial challenges due to overlaps with TB and inter-observer variation. In this research, a new deep learning approach is introduced for automatic detection of NTM with 2D slice representations extracted from DICOM volumes containing only the lung region. With a dataset of 1301 CT scans where 430 were NTM and 871&#xa0;TB, a classifier is created for automatic discrimination between NTM and non-NTM cases. The suggested architecture utilizes the combination of CNN backbone for local features extraction and Transformer encoder for global context modeling, which makes possible the learning of complex multi-scale pathological features. Automatic extraction of the ROI from the lungs allows for enhanced image quality through noise filtering and retention of morphological features. The experiments show that the suggested method demonstrates consistent and significant improvement in comparison with baseline models in terms of diagnostic accuracy, which indicates promising opportunities of application for the automation of NTM detection process.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Hybrid CNN–Transformer Deep Learning Framework for Automated Detection of Nontuberculous Mycobacterial Pulmonary Disease Using Lung-Cropped 2D HRCT Representations

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Ch.Satya Keerthi.N.V.L,
  • P. Siddaiah

摘要

Nontuberculous mycobacterial pulmonary disease (NTM-PD) is a progressive infectious illness with various radiological appearances such as bronchiectasis, tree-in-bud opacity, nodules, and parenchymal distortion. Manual diagnosis with 2D HRCT has been associated with substantial challenges due to overlaps with TB and inter-observer variation. In this research, a new deep learning approach is introduced for automatic detection of NTM with 2D slice representations extracted from DICOM volumes containing only the lung region. With a dataset of 1301 CT scans where 430 were NTM and 871 TB, a classifier is created for automatic discrimination between NTM and non-NTM cases. The suggested architecture utilizes the combination of CNN backbone for local features extraction and Transformer encoder for global context modeling, which makes possible the learning of complex multi-scale pathological features. Automatic extraction of the ROI from the lungs allows for enhanced image quality through noise filtering and retention of morphological features. The experiments show that the suggested method demonstrates consistent and significant improvement in comparison with baseline models in terms of diagnostic accuracy, which indicates promising opportunities of application for the automation of NTM detection process.