Vision Transformer Based Detection Of Chronic Pulmonary Aspergillosis Lung Infections In Chest X-Ray Images

Chronic pulmonary aspergillosis (CPA) is a lung infection caused by the fungus Aspergillus. The infection can occur when the immune system is compromised or there is damage to the lung tissue. Due to its rarity, physicians often do not consider CPA as an initial diagnosis and may not perform relevant tests, leading to a lack of targeted treatment in the early stages of the disease and potentially affecting the effectiveness of subsequent treatments. This paper employs a model based on the Vision Transformer (ViT) and utilizes the self-supervised DINO framework to train a classifier for determining whether a given chest X-ray (CXR) image belongs to a patient with CPA. To the best of our knowledge, there is currently no deep-learning based study for the diagnosis of CPA using CXR images. We demonstrate the effectiveness of ViT trained with the DINO framework for classifying CXR images using both in-house and public datasets. For results in three-class classification (normal, CPA, non-CPA-abnormal), we can achieve macro F1 of 0.768. For binary classification (normal vs. CPA), we obtain macro F1 of 0.935, with recall of 0.803 and precision of 0.953 for the CPA class.