0% found this document useful (0 votes)
16 views24 pages

Deep Learning for Brain Tumor Detection

The document presents a deep learning pipeline for brain tumor detection using U-Net for segmentation and Xception for classification, addressing the challenge of class imbalance through LoRA-driven synthetic data augmentation. The methodology includes detailed steps for data preparation, model training, and evaluation, demonstrating significant improvements in segmentation and classification performance. The study concludes with the development of a user-friendly interface for practical application and outlines future work directions for enhancing the system.

Uploaded by

Prottoy
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
16 views24 pages

Deep Learning for Brain Tumor Detection

The document presents a deep learning pipeline for brain tumor detection using U-Net for segmentation and Xception for classification, addressing the challenge of class imbalance through LoRA-driven synthetic data augmentation. The methodology includes detailed steps for data preparation, model training, and evaluation, demonstrating significant improvements in segmentation and classification performance. The study concludes with the development of a user-friendly interface for practical application and outlines future work directions for enhancing the system.

Uploaded by

Prottoy
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd

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?

You might also like