Optimizing Tumor Detection in Brain MRI with One-Class SVM and Convolutional Neural Network-Based Feature Extraction.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Neuro-oncology

Background:

• Early brain tumor detection is crucial for patient outcomes.
• Medical imaging datasets often suffer from class imbalance, hindering traditional AI classification.
• Developing robust AI for anomaly detection in scarce pathological data is a significant challenge.

Purpose of the Study:

• To investigate the efficacy of One-Class Support Vector Machine (OCSVM) combined with deep learning feature extraction for brain tumor anomaly detection.
• To compare the performance of various deep learning architectures (DenseNet121, VGG16, MobileNetV2, InceptionV3, ResNet50) and classical methods for feature extraction.
• To evaluate a pure Convolutional Neural Network (CNN) approach for direct classification without OCSVM.

Main Methods:

• Extracted features from healthy brain MRI images using deep learning architectures and classical techniques.
• Trained a One-Class Support Vector Machine (OCSVM) exclusively on features from healthy brain images.
• Compared the performance of hybrid CNN-OCSVM models against a pure CNN classification model.

Main Results:

• Hybrid CNN-OCSVM models significantly improved anomaly detection over handcrafted features.
• DenseNet121 (94.83% accuracy) and VGG16 (95.33% accuracy) showed strong performance in hybrid models.
• A pure CNN model achieved superior accuracy (97.83%), demonstrating effective direct feature learning from MRI data.
• MobileNetV2 offered a balance between accuracy and computational efficiency.

Conclusions:

• AI models, particularly pure CNNs, can reliably detect brain tumor anomalies in imbalanced MRI datasets without pathological labels.
• This approach offers a promising solution for clinical settings with limited abnormal samples.
• Future work includes optimizing inference time, dataset expansion, and enhancing model explainability for clinical trust.