Detection of Thymoma Disease Using mRMR Feature Selection and Transformer Models.

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Thymoma is a rare malignant tumor originating in the thymus gland, primarily affecting adults.
• Current diagnostic methods rely on expert opinion, with a growing need for advanced technologies like AI.
• Early detection systems are increasingly incorporating sophisticated models, including transformer models.

Purpose of the Study:

• To evaluate the efficacy of transformer models in detecting thymoma from medical images.
• To compare the performance of transformer models against traditional deep learning approaches for thymoma diagnosis.
• To develop an efficient and accurate AI-driven system for thymoma detection.

Main Methods:

• Utilized a dataset of thymoma and non-thymoma images from Fırat University.
• Employed preprocessing techniques including region of interest (ROI) cropping.
• Trained four transformer models (Deit3, Maxvit, Swin, ViT), fused features from the best performers, and applied mRMR feature selection with SVM classification.

Main Results:

• Achieved 100% overall accuracy in thymoma detection using mRMR feature selection with a reduced feature set.
• Cross-validation confirmed the robustness of the approach, yielding 99.22% overall accuracy.
• The combined feature set from Deit3 and ViT models demonstrated superior performance.

Conclusions:

• The proposed AI approach, leveraging transformer models, significantly enhances thymoma detection capabilities.
• This study highlights the potential of advanced AI techniques in improving diagnostic accuracy for rare cancers.
• The findings underscore the added value of this method for clinical application in thymoma diagnosis.