RadioGuide-DCN: A Radiomics-Guided Decorrelated Network for Medical Image Classification

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Deep learning methods in medical imaging face challenges like overfitting due to limited datasets.
• Traditional radiomics approaches often lack specificity and fail to capture complex pathological details.
• Integrating diverse imaging modalities (radiography, ultrasound, CT, MRI) is crucial for comprehensive diagnosis.

Purpose of the Study:

• To develop an innovative radiomics-guided decorrelated classification network (RadioGuide-DCN) for improved medical image analysis.
• To address limitations of existing deep learning and traditional radiomics methods in capturing complex pathological information.
• To enhance the model's ability to discern local details and global patterns in medical images.

Main Methods:

• Proposed RadioGuide-DCN integrates radiomics features as prior information into deep neural networks.
• Employed a feature decorrelation loss mechanism and an anti-attention feature fusion module to reduce redundancy.
• Utilized a Kolmogorov-Arnold Network (KAN) classifier with learnable activation functions for boosted performance.

Main Results:

• RadioGuide-DCN achieved 93.63% accuracy in BUSI image classification.
• The method consistently outperformed conventional radiomics and deep learning approaches in various medical imaging tasks.
• Demonstrated significant improvements in classification accuracy and Area Under the Curve (AUC) scores.

Conclusions:

• RadioGuide-DCN offers a novel paradigm for integrating deep learning with traditional imaging analysis.
• The proposed method shows broad clinical application potential, especially in tumor detection and disease diagnosis.
• This approach enhances the capacity to capture both local and global patterns, leading to more accurate medical image classification.