AI Research Analysis

Deep Learning Architectures Explored

The study compared several prominent CNN architectures, each with distinct structural characteristics, known for their success in medical image classification:

• ResNet-50: A deep residual network, known for mitigating vanishing gradients and achieving high accuracy.
• VGG-16: A classic architecture with a uniform and deep structure of 3x3 convolutional layers.
• DenseNet-121: Features dense connections where each layer receives inputs from all preceding layers, enhancing feature reuse.
• InceptionV3: Utilizes inception modules to capture features at multiple scales, reducing computational cost.
• MobileNet: A lightweight model optimized for mobile and embedded vision applications, leveraging depthwise separable convolutions for efficiency.

These models were chosen to cover a spectrum from deeper, more complex networks to lightweight, efficient ones, enabling a comprehensive performance comparison.

Rigorous Experimental Methodology

The experimental setup focused on ensuring fair comparison and reproducibility:

• Dataset: A curated dataset of ~7,023 2D T1-weighted contrast-enhanced MRI slices from Figshare, SARTAJ, and Br35H, categorized into glioma, meningioma, pituitary, and no-tumor. Mislabeled glioma samples were corrected for reliability.
• Data Splitting: Patient-wise stratified 70%/15%/15% train/validation/test splits were enforced to avoid slice-level leakage and ensure generalization.
• Preprocessing & Augmentation: Images were resized to 224x224, min-max normalized to [0,1], and enhanced with CLAHE. Training data was augmented with random rotations, flips, shifts, zoom, and Gaussian noise.
• Training: Models were initialized with ImageNet pretrained weights (transfer learning), fine-tuned using Adam optimizer (LR=1e-3, batch size=32, max 30 epochs), and employed early stopping (patience=5) and ReduceLROnPlateau (patience=3).
• Evaluation: Performance was measured using accuracy, precision, recall, F1-score, and AUC, with confusion matrices and Grad-CAM visualizations for interpretability.

Key Performance Findings

The comparative analysis revealed distinct strengths, with MobileNet consistently outperforming others in terms of overall accuracy and computational efficiency.

• MobileNet: Achieved the highest accuracy of 96.6%, demonstrating its strong capability as a lightweight CNN for reliable brain-tumor classification. Its depthwise separable convolutions were key to its efficiency.
• DenseNet-121: Followed MobileNet with 95.2% accuracy.
• ResNet-50: Performed moderately with 94.8% accuracy.
• InceptionV3: Achieved 93.7% accuracy.
• VGG-16: Showed comparable performance to ResNet-50 with 93.3% accuracy.

These results highlight that lightweight architectures can achieve superior performance, making them practical for resource-constrained clinical settings. The study provides insights into selecting suitable CNN models for automated brain tumor detection.