SVD-CLAHE

SVD-CLAHE enhances contrast and balances class distributions in chest X-ray (CXR) datasets to improve deep-learning-based COVID-19 detection.

Key Features:

• SVD-CLAHE Boosting Methodology: Applies Singular Value Decomposition (SVD)-based contrast enhancement combined with Contrast Limited Adaptive Histogram Equalization (CLAHE) to augment underrepresented classes.
• Random Under Sampling (RUS): Reduces samples from overrepresented classes via Random Under Sampling to achieve a more equitable class distribution.
• Balanced Weighted Categorical Cross Entropy (BWCCE): Uses a loss function that dynamically adjusts category weights during training to mitigate class imbalance.
• Model architectures and pretraining: Validates the framework using ResNet-50 and VGG-19 architectures pre-trained on the ImageNet dataset.
• Performance metrics and benchmarking: Reports 95% F1 score, 94% accuracy, 95% recall, 96% precision, and 96% AUC and compares performance against InceptionV3, DenseNet-121, Xception, Covid-Lite, and Covid-Net.

Scientific Applications:

• COVID-19 detection from CXR images: Enables automatic detection of COVID-19 in chest X-ray images to support early diagnosis.
• Class imbalance mitigation in medical imaging: Improves deep learning model performance on minority classes through targeted augmentation and loss weighting.

Methodology:

Data augmentation uses SVD-based contrast enhancement and CLAHE for oversampling minority classes and Random Under Sampling for major classes; model validation employs ResNet-50 and VGG-19 pre-trained on ImageNet, with performance reported using F1 score, accuracy, recall, precision, and AUC.