Deep Learning for Brain Tumor Detection:
U-Net Segmentation and Xception
Classification using Low Rank Adaptation -
Driven Synthetic Data Augmentation
Md. Muhaimenul Haque Prottoy
Supervisor: Anika Anjum
Bachelor of Science in Biomedical Engineering
Department of Biomedical Engineering
Jashore University of Science and Technology
April 2025
Outline
Introduction & Motivation
Problem Statement & Objectives
Background Concepts
Methodology
Overall Pipeline
Datasets
U-Net Segmentation
Synthetic Data Generation (LoRA + Flux)
Xception Classification
Baseline Comparisons
Implementation Details
Results & Discussion
Segmentation Performance
Synthetic Image Quality
Classification Performance (Imbalance & Augmentation Impact)
Comparative Analysis
Conclusion & Future Work
Q&A
Introduction & Motivation
Challenge: Brain tumor detection is critical but complex.
MRI: Primary diagnostic tool, but manual interpretation is:
Time-consuming
Subjective & expertise-dependent
Prone to errors
Deep Learning Potential: Offers automation, improved accuracy, and
efficiency for:
Tumor Segmentation (Isolating tumor)
Tumor Classification (Identifying tumor type)
(Image: Add example MRI scan manually)
Problem Statement
Need for Automated Tools: Improve diagnostic speed and reliability.
Key Challenge: Class Imbalance
Medical datasets often have uneven class distributions.
Models trained on imbalanced data become biased.
Leads to poor performance on underrepresented classes.
Goal: Develop a robust pipeline addressing segmentation/classification
AND class imbalance.
Objectives
Develop a deep learning pipeline for brain tumor detection
(segmentation + classification).
Implement U-Net for accurate tumor segmentation.
Implement Xception for multi-class tumor classification.
Address class imbalance using LoRA-driven synthetic data
augmentation.
Evaluate the impact of synthetic data on performance and confidence.
Compare LoRA-based augmentation against traditional methods.
Deploy the system via a user-friendly interface (Gradio).
Background - Key Concepts
MRI: Detailed brain structure visualization.
Deep Learning: Neural networks learning complex patterns.
CNNs: Effective for image analysis.
U-Net: Specialized CNN for biomedical segmentation.
Xception: Efficient CNN for image classification.
Synthetic Data Augmentation: Creating artificial data for balance.
LoRA: Efficiently fine-tunes large models.
Flux 0.1 Dev: Text-to-image model for synthetic MRI generation.
Background - U-Net & Xception
U-Net:
Purpose: Segmentation (Pixel-level classification).
Architecture: Encoder -> Decoder + Skip Connections.
(Diagram: Add Figure 2.1 simplified manually)
Xception:
Purpose: Classification (Image label assignment).
Architecture: Depthwise Separable Convolutions, pre-trained.
(Diagram: Add Figure 2.2 simplified manually)
Methodology - Overall Pipeline
(Flowchart: Add Figure 3.1 manually)
Input: Brain MRI Scan
Step 1: U-Net Segmentation (Predict Mask)
Step 2: Classification Preprocessing
Step 3: Handle Imbalance (Offline: LoRA + Flux)
Train class-specific LoRA.
Generate synthetic images.
Merge synthetic & real data.
Step 4: Xception Classification (Train on balanced data)
Step 5: Evaluation (Metrics)
Step 6: Deployment (Gradio Interface)
Methodology - Datasets
Segmentation Dataset (Kaggle Ref [35])
3,064 MRI images + 3,064 masks.
Used for U-Net training.
Preprocessing: Resize (256x256), Normalize.
Classification Dataset (Kaggle Ref [36])
~7,023 MRI images (4 classes).
Split: ~5712 Train / ~1311 Test.
Training Subset Used: Balanced 500 images/class (2000 total).
Preprocessing: Resize (299x299), Normalize, RGB.
Used for Xception training & augmentation experiments.
Methodology - U-Net Segmentation
Model: U-Net Architecture (Fig 3.2)
Input: 256x256 RGB (from grayscale)
Loss Function: Dice Loss
Optimizer: Adam (LR=0.0001)
Training: 50 Epochs, BS=16, Early Stopping
Evaluation Metrics: Dice Score, IoU
Methodology - Synthetic Data
Generation (LoRA + Flux)
Goal: Balance classification dataset for underrepresented class.
Process:
1. Simulate Imbalance: Reduce one class (e.g., 500 -> 250).
2. Train LoRA: Fine-tune on 100 images of reduced class ([Link]).
3. Generate Images: Use LoRA + Flux + Text Prompts (Hugging Face).
4. Balance Dataset: Combine real (reduced) + synthetic images.
(Flowchart: Add Figure 3.4 manually)
Methodology - Xception Classification
& Baselines
Model: Xception (pre-trained) + Custom Layers
Input: 299x299 RGB
Loss: Categorical Crossentropy
Optimizer: Adam (LR=0.0001)
Training: 50 Epochs, BS=32, Early Stopping
Iterative Training Scenarios:
Original Balanced Dataset
Reduced Datasets (Baseline drop)
LoRA-Augmented Datasets (Proposed)
Baseline Augmented Datasets:
Random Augmentation
Class Weighting
Oversampling
Evaluation: Accuracy, Precision, Recall, F1, Confusion Matrix
(Flowchart: Add Figure 3.6 manually)
Implementation Details
Platforms:
Google Colab: U-Net & Xception Training (T4 GPU)
[Link]: LoRA Model Training
Hugging Face Spaces: Synthesis, Hosting, Gradio
Libraries:
TensorFlow/Keras, PyTorch
OpenCV, NumPy, Gradio
(Logos: Add manually if desired)
Results - Segmentation Performance
Metrics:
Average Dice Score: 88.43%
Average IoU: 84.21%
Interpretation: High overlap, effective segmentation.
Visual Results:
(Show Figure 6.1 manually - examples needed)
Results - Synthetic Image Quality
Visual Comparison:
(Show Figure 6.2 manually - Real vs. Synthetic examples)
Observation: Synthetic images capture key features, suitable for
augmentation.
Results - Impact of Class Reduction &
LoRA
Class Reduction Impact:
Significant drop in metrics (Acc, P, R, F1) when class reduced.
(Show Graphs like Fig 6.4, 6.10, 6.16, 6.22 manually)
LoRA Augmentation Impact:
Recovery and improvement over reduced baseline.
(Show Graphs like Fig 6.5, 6.11, 6.17, 6.23 manually)
Avg Improvement (LoRA vs Reduced): Acc +2.55%, P +2.38%, R
+2.37%, F1 +2.51%
Results - Confidence Score
Improvement
(Show Figure 6.3 manually - Confidence meter plots)
Observation: Model confidence increased significantly after LoRA
augmentation, especially for underrepresented classes.
Results - Comparative Analysis of
Augmentation
(Show Summary Table/Graphs manually - Based on Table 6.1 etc.)
Comparison Metrics: Accuracy, Precision, Recall, F1-Score
Techniques Compared:
Original / Reduced (Baseline)
LoRA Augmentation (Proposed)
Random Augmentation
Class Weighting
Oversampling
Key Finding: LoRA consistently outperformed baseline techniques.
Results - Confusion Matrices
(Show selected matrices manually: Fig 6.28, 6.29, 6.33)
Observations:
Original: Good performance.
Reduced: Increased misclassifications.
LoRA Balanced: Reduced misclassifications, stronger diagonal.
Results - Gradio Interface
(Add Screenshot/Description manually)
Functionality: Upload MRI -> Get Segmentation Mask & Classification +
Confidence
Purpose: Demonstrates practical application.
Discussion
U-Net: Effective for segmentation (High Dice/IoU).
Xception: Good for classification, especially with balanced data.
Class Imbalance: Clearly hurts performance.
LoRA Augmentation Success:
Generated realistic synthetic data.
Mitigated class imbalance effectively.
Outperformed traditional methods.
Boosted metrics AND model confidence.
Pipeline Integration: Successful combination enhanced by augmentation.
Conclusion
Developed end-to-end pipeline (U-Net + Xception).
LoRA-driven synthetic data augmentation is highly effective against
class imbalance.
Achieved significant improvements in metrics and confidence vs.
baselines.
Provides a more robust and reliable approach for automated MRI
analysis.
Deployed practical interface (Gradio).
Future Work
Advanced Architectures: ViTs, Hybrids.
Improved Synthesis: Diffusion Models, cGANs.
Dataset Expansion: Multi-centric, Multi-modal, Longitudinal.
Clinical Integration: Real-world testing, UI refinement.
Integrate Patient Data: Clinical, Genetic.
Continuous Learning: Model updates.
Thank You & Q&A
Acknowledgements
Thank you for your attention.
Questions?