Image Recognition in Chest Radiographs for Chronic Pulmonary Aspergillosis Detection through Deep Learning Process

<bold>Background:</bold> Chronic Pulmonary Aspergillosis (CPA) is a chronic Aspergillus infection in immunocompetent hosts. Radiological finding of CPA have distinct features, including cavities, pleural thickening, aspergilloma, and fibrotic lung. Despite of these radiographic signatures, there exists no established alternative intelligence (AI) model leveraging for CPA detection. This study aimed to developing an AI model tailored for CPA identification. Method The study harnesses a deep learning framework incorporating the Vision Transformer (ViT) as its backbone, augmented with the self-supervised DINO framework and Synthetic Minority Over-sampling Technique (SMOTE). Retrospective data collection involves patients who were laboratory confirmed cases presenting with abnormal chest X-rays (CXR) indicative of CPA from the institutional dataset (TVGH dataset). Additionally, the training dataset encompasses a roentgenographic repository sourced from the National Institutes of Health (NIH) labled with other pulmonary abnormalities (non-CPA) and normal CXRs. Cross validation tests entail partitioning 80% of the combined TVGH and NIH datasets for training and the remaining 20% for independent testing, iterated five times. <bold>Results:</bold> A total of 275 CXRs from 79 patients with CPA, 11011 CXRs from patients with non-CPA and 10,000 normal CXRs were obtained. Considering the CXRs including non-CPA cases (3 categories: CPA, non-CPA, normal) or not (2 catefories: CPA, normal), the resultant model exhibits a sensitivity of 96%-97.8% and specificity of 100-99%% in CPA detection, accompanied by an impressive F1 score of 0.96-0.98 in model processing using 3 and 2 categories datasets, respectively.