Comparative Analysis of Brain Tumor Detection
Using CNNs, U-Net Architecture and Transfer
Learning
Presenter: Vistrit Kumar Rai.
SCSE, Galgotias University
Conference: [ICAISI]
Date: 31/05/2025
Abstract
• Overview of brain tumor detection
importance.
• Techniques compared: CNN, U-Net, ResNet-
50.
• ResNet-50 outperforms in accuracy and
segmentation.
Introduction
• Importance of early tumor detection.
• Common imaging: MRI, CT, PET.
• Deep learning in medical imaging.
Methods Compared
• CNN: Efficient but less accurate.
• U-Net: Balanced performance.
• Transfer Learning (ResNet-50): High accuracy,
high cost.
Dataset Description
• BraTS Challenge Dataset.
• MRI Types: T1, T2, FLAIR.
• Tumor regions: ET, ED, NCR/NET.
Preprocessing Techniques
• Normalization, resizing, data augmentation.
• Image registration for consistency.
Model Architectures
• CNN: Basic layers.
• U-Net: Encoder-decoder with skips.
• ResNet-50: Fine-tuned from ImageNet.
Training and Evaluation
• Train/Validation/Test split.
• Optimizer: Adam.
• Metrics: Accuracy, DSC, IoU, etc.
Comparative Performance
• Accuracy: CNN(89.3%), U-Net(94.5%), ResNet-
50(96.1%)
• DSC: CNN(0.72), U-Net(0.85), ResNet-50(0.89)
• IoU: CNN(65.4%), U-Net(78.3%), ResNet-
50(81.5%)
Computational Efficiency
• Training Time: CNN(8h), U-Net(12h), ResNet-
50(16h)
• Inference Time: CNN(0.34s), U-Net(0.43s),
ResNet-50(0.52s)
• GPU Usage: CNN(220MB), U-Net(280MB),
ResNet-50(360MB)
Error Analysis
• CNN: Fast, poor segmentation.
• U-Net: Good accuracy, noise sensitive.
• ResNet-50: Best accuracy, high resources.
Conclusion
• CNN: Low-resource fit.
• U-Net: Balanced option.
• ResNet-50: Best overall with resource
availability.
References
• Key papers and tools:
• - Géron, O'Reilly Media
• - Esteva et al., Nature
• - BraTS Challenge Dataset
Q&A
• Thank you!
• Questions?

Brain_Tumor_Detection_Presentation.for project