Keros-Net: a convolutional block attention module-squeeze-and-excitation-integrated hybrid learning framework for...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Olfactory fossa depth classification is crucial for safe endoscopic sinus surgery.
• Current methods can have observer variability.
• Accurate preoperative assessment prevents surgical complications like cribriform plate injury.

Purpose of the Study:

• Develop a hybrid deep learning (DL) and machine learning (ML) framework.
• Automate olfactory fossa depth classification on paranasal CT images (Keros classification).
• Improve accuracy, reduce observer variability, and support safer surgeries.

Main Methods:

• Retrospective analysis of 481 paranasal CT scans (1,549 slices).
• Deep features extracted using CBAM-SE enhanced DenseNet architectures.
• Feature selection via RFE, PCA, SelectKBest, and SHAP.
• Classification using SVM, Random Forest, XGBoost, Logistic Regression, and Naive Bayes.
• Five-fold cross-validation for performance assessment.

Main Results:

• Baseline DenseNet169 achieved 88.37% accuracy.
• Hybrid approach with RFE + SVM/Logistic Regression reached 97.90% accuracy.
• Recursive feature elimination (RFE) was the most effective feature selection method.
• SVM provided consistently balanced classification results.

Conclusions:

• A hybrid framework integrating DL (DenseNet with CBAM-SE) and ML classifiers with optimized feature selection offers highly accurate Keros type classification.
• This approach surpasses conventional DL models, providing a robust tool for objective radiological assessment.
• The system enhances surgical planning, reduces operator dependency, and increases patient safety.