ALCNN: Attention based lightweight convolutional neural network for pneumothorax detection in chest X-rays. (January 2023)

• Record Type: Journal Article
• Authors: Agrawal, Tarun; Choudhary, Prakash
• Is Part Of: Biomedical signal processing and control. Volume 79(2023)Part 1
• Publication Date: 2023-01
• Subjects: Pneumothorax detection -- Deep learning -- Chest X-ray -- Deep convolutional neural network; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104126
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Pneumothorax can be life-threatening if not diagnosed on time. Most of the research studies for pneumothorax detection in chest X-rays (CXR) images have applied the transfer learning approach. In this study, we have explored the parameter-efficient attention based lightweight convolutional neural network (ALCNN) for pneumothorax detection in CXR images. It has stacked convolutional layers with the attention mechanism to re-calibrate channel-wise feature maps. Furthermore, we have also evaluated three different transfer learning approaches to compare the performance with ALCNN. The ALCNN has obtained the comparable results with 10x fewer parameters in comparison to VGG-19 and ResNet-50 architectures. The ALCNN achieved a 0.73 sensitivity and an area under curve (AUC) of 0.79 on the held-out test set. While pre-trained and fine-tuned, VGG, 19, achieved a sensitivity value of 0.52 and 0.71 and an AUC value of 0.72 and 0.81, respectively. Similarly, pre-trained and fine-tuned ResNet-50 achieved a sensitivity value of 0.69 and 0.74 and an AUC value of 0.77 and 0.80, respectively. In the third transfer learning approach, VGG, 19, and ResNet-50 obtained sensitivity values of 0.65 and 0.69, respectively, while the AUC value of 0.79 for both VGG-19 and ResNet-50. The results obtained in our study show that transfer learning did not have a significant impact on performance for pneumothorax detection in CXR. Graphical abstract: Highlights: Parameter-efficient ALCNN model is proposed for pneumothorax detection in chest X-rays. The ALCNN obtained comparable results with ten times less parameters when compared with three transfer learning approaches on two deep architectures. Obtained results show that the transfer learning approaches did not have a significant effect on the performance for pneumothorax detection.