Attention enhanced residual network for automatic pulmonary tuberculosis detection on chest radiographs images

• Published in: Digital signal processing 2025-04, Vol.159, p.104975, Article 104975
• Main Authors: Liu, Yulong, Liu, Feng, Tu, Shixin, Liu, Shangqing, Han, Baoru
• Format: Article
• Language: English
• Subjects: Attention mechanisms; Convolutional neural networks; Tuberculosis detection; Warmup-to-finetuning
• ISSN: 1051-2004
• DOI: 10.1016/j.dsp.2024.104975

Pulmonary tuberculosis (TB) is a common disease in developing countries that spreads through direct contact or through the air. Early detection of people with tuberculosis plays a key role in preventing transmission and saving costs ultimately. Chest X-ray scan is a recommended screening technique to spot the pulmonary abnormalities. In this study, we introduce a classification approach aiming to improve the classification accuracy of TB with chest X-ray images. Specifically, we introduce the CBAM module to ResNet (CBAM-ResNet) exploiting both the spatial and channel information of features maps. Considering the limited data for training, we propose a two-stage warmup-to-finetuning (W2F) training strategy that warming up the parameters with simple start and then take fully advantage of data augmentation techniques such as mix-up to fine-tune the parameter. Furthermore, we leverage the stochastic gradient descent (SGD) equipped with gradient centralization (SGD-GC) and dynamic learning rate as the optimizer to optimize the parameters. The proposed two-stage TB detection method was evaluated using two public datasets. The experimental results indicate that our method outperformed existing state-of-the-art CNN-based methods, achieving an average test accuracy of 91.1% on the Shenzhen dataset and 86.4% on the Montgomery dataset. Overall, the proposed method may aid doctors and radiologists in early detection of TB, potentially increasing patient survival rates. •CBAM-enhanced ResNet integrates spatial and channel attention, refining feature focus and minimizing irrelevant data's influence in training.•Proposed a two-stage W2F strategy, combining warmup with mixup-based augmentation to fine-tune network parameters on limited data.•Utilized SGD with gradient centralization (SGD-GC) and dynamic learning rate to effectively optimize network parameters.