ABSTRACT
Aim/Background
In this work, a novel method for improving the quality of healthcare in the diagnosis of Interstitial Lung Disease (ILD) using High-Resolution Computed Tomography (HRCT) images is proposed.
Materials and Methods
In contrast to previous research that necessitated the human identification of Regions of Interest (ROI), a two-phase deep learning method is presented. First, multi-scale feature extraction is used to precisely segment the lung in HRCT images using a conditional Generative Adversarial Network (c-GAN). A Support Vector Machine (SVM) classifier classifies the characteristics extracted by a pretrained ResNet50 from the segmented lung image into seven ILD classes in the second step.
Results
The two-step approach that is being offered improves efficiency by doing away with the necessity for ROI extraction. The superiority of the method is demonstrated by performance comparison with patch-based and whole-image-based algorithms. The suggested method reduces false alarms by achieving a maximum classification accuracy of 94.65% for the normal class. Despite having the lowest accuracy (84.12%), the consolidation class performs better than other whole-image-based methods.
Conclusion
The suggested two-stages ILD classifier performs much better due to the step-by-step improvement in the deep learning method. This work lays the groundwork for advanced decision support systems in the pharmaceutical industry and advances pharmaceutical research and education. The method proposed improves knowledge of the pathophysiology of ILD and allows for customized treatment approaches.