Detection of Aspergilloma Disease Using Feature-Selection-Based Vision Transformers.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Fungal Infections

Background:

• Aspergilloma disease, caused by Aspergillus fungus, forms fungal masses in organs like lungs and sinuses.
• Diagnosis relies on expert opinion and advanced technologies, with deep learning models aiding early detection.
• Current treatments involve surgical methods, emphasizing the need for precise diagnostic tools.

Purpose of the Study:

• To evaluate the efficacy of vision transformers (ViTs) for aspergilloma disease detection.
• To develop an advanced deep learning framework for improved diagnostic accuracy.
• To compare ViT performance against traditional deep learning models in identifying aspergilloma.

Main Methods:

• Utilized a dataset of aspergilloma and non-aspergilloma images from thoracic surgery patients.
• Employed pre-processing, data augmentation, and three ViT models (vit_base_patch16, vit_large_patch16, vit_base_resnet50) for training.
• Integrated feature selection (Chi2, mRMR, Relief) and fusion techniques followed by Support Vector Machines (SVM) classification.

Main Results:

• The proposed method, utilizing SVM classification, achieved an outstanding 99.70% overall accuracy.
• Cross-validation confirmed the high performance and reliability of the diagnostic approach.
• Vision transformers demonstrated significant potential in enhancing aspergilloma detection accuracy.

Conclusions:

• The developed method, leveraging ViTs and SVM, offers a highly effective tool for aspergilloma diagnosis.
• This approach shows promise for integration into clinical settings for earlier and more accurate disease identification.
• The study underscores the advancement of AI in medical diagnostics, particularly for challenging fungal infections.