• Research Overview
• The Novel Contribution — Bridge Transfer Learning
• Published Results
• Model Architecture
• Datasets
• Web Applications
• Project Structure
• Setup & Running
• Authors

This project proposes a Modality-Bridge Transfer Learning framework for medical CT scan image classification. The core challenge addressed is the domain mismatch between natural image pre-training (ImageNet) and medical image target domains.

Traditional transfer learning:

ImageNet (natural photos) ──────────────────→ Lung CT scans
 Source Target
 [Large domain gap — different textures, structures, imaging artifacts]

ImageNet → Kidney CT scans → Lung CT scans
 Source Bridge Target
 [Same CT modality] [Same CT modality]
 [Smaller gap] [Organ-level feature transfer]

By inserting a bridge domain — kidney CT scans, which share the same acquisition modality (CT) as the target lung CT — the model progressively adapts from natural image features to medical CT features before reaching the target task. This reduces distribution mismatch and improves generalization even with limited labelled target data.

┌─────────────────────────────────────────────────────────────────────────────┐
│ Modality-Bridge Transfer Learning Pipeline │
│ │
│ Step 1: Pre-train on ImageNet (natural images — 14M images, 1000 classes) │
│ │ │
│ ▼ │
│ Step 2: Fine-tune on Kidney CT Dataset (bridge domain — same CT modality) │
│ → VGG19 learns CT-specific features: tissue density, organ │
│ boundaries, contrast gradients, bone structure │
│ → Achieves 95.6% accuracy on kidney tumor classification │
│ │ │
│ ▼ │
│ Step 3: Transfer to Lung CT Dataset (target domain) │
│ → Model already understands CT imaging characteristics │
│ → Fine-tune final layers for lung cancer histology │
│ → Achieves 93% accuracy (vs 90% without bridge) │
│ │
│ KEY INSIGHT: Same modality (CT) → Shared low-level features │
│ (edges, textures, density gradients) transfer cleanly across organs │
└─────────────────────────────────────────────────────────────────────────────┘

The paper also introduces Multi-Bridge Transfer Learning, which allows leveraging multiple source domains simultaneously, further improving robustness and handling data imbalance.

Key finding: Bridge transfer learning improves Lung CT classification by +3% over the standalone baseline — demonstrating effective cross-organ knowledge transfer.

Conclusion from Table II: CT-to-CT transfer (93%) significantly outperforms USG-to-CT transfer (75%), confirming that same-modality bridging is critical to the method's effectiveness.

Note on discrepancy: The paper reports 90% for the lung CT baseline — this corresponds to the training accuracy of the balanced dataset model. The test accuracy of 84.4% reflects true generalisation on unseen data. The bridge learning result of 93% (paper) represents the improvement achieved by the cross-organ knowledge transfer.

Input Image (224×224×3)
 │
 ▼
┌─────────────────────┐
│ Block 1 │ 2× Conv(64, 3×3) + ReLU → MaxPool(2×2)
├─────────────────────┤
│ Block 2 │ 2× Conv(128, 3×3) + ReLU → MaxPool(2×2)
├─────────────────────┤
│ Block 3 │ 4× Conv(256, 3×3) + ReLU → MaxPool(2×2)
├─────────────────────┤
│ Block 4 │ 4× Conv(512, 3×3) + ReLU → MaxPool(2×2)
├─────────────────────┤
│ Block 5 │ 4× Conv(512, 3×3) + ReLU → MaxPool(2×2)
├─────────────────────┤
│ Flatten │ 7×7×512 = 25,088 features
├─────────────────────┤
│ FC-4096 + Dropout │ Fine-tuned for medical domain
├─────────────────────┤
│ FC-4096 + Dropout │ Fine-tuned for medical domain
├─────────────────────┤
│ Output (Softmax) │ Kidney: 2 classes | Lung: 4 classes
└─────────────────────┘

• Proven feature extractor on ImageNet (top-5 accuracy: 92.7% on 1000 classes)
• Deep enough to learn hierarchical medical image features (edges → textures → organ structures → pathology)
• Simple uniform architecture makes fine-tuning predictable and controllable
• Well-studied transfer learning properties

• Task: Binary classification — Normal vs Tumor
• Classes: Normal, Tumor
• Modality: CT scan (KUB — Kidney, Ureter, Bladder)
• Role: Bridge domain — same CT modality as target

• Task: Multi-class classification — 4 histological subtypes
• Classes:
• Adenocarcinoma — solid nodule with spiculated margins, ground-glass opacity
• Large Cell Carcinoma — large mass with irregular borders and necrotic center
• Squamous Cell Carcinoma — cavitary lesion with thick walls and calcifications
• Normal — healthy lung tissue
• Modality: Chest CT scan

• Resize all images to 224×224×3 (VGG19 input requirement)
• Normalize pixel values to [0, 1]
• Data augmentation: shift, scale, rotate, flip, brightness adjustment, noise introduction
• Balanced dataset created to address class imbalance (original dataset overrepresented normal class)

Two Flask web apps are included for live inference:

• Upload CT scan image (.jpg, .jpeg, .png) or provide a URL
• VGG19 inference with top-2/top-3 class probabilities
• Clinical recommendation text per prediction
• Responsive web interface

Tumor Classification/
│
├── Notebooks (Training)
│ ├── KUB-ct-scan-VGG19(for-kidney-tumors).ipynb # Kidney tumor VGG19 training
│ ├── chest-ct-scan-VGG19-with-transfer-learning.ipynb # Lung cancer VGG19 training
│ ├── lung-ct-scan-VGG19(for-lung-tumors)-D2.ipynb # Lung cancer balanced training
│ └── Comparsion.ipynb # Model comparison analysis
│
├── Trained Models
│ ├── kidney_tumor_model.hdf5 # Kidney VGG19 model weights
│ ├── lung_cancer_model.hdf5 # Lung VGG19 model weights (unbalanced)
│ ├── lung_cancer_B_model.hdf5 # Lung VGG19 model weights (balanced)
│ └── chest_CT_SCAN.h5 # Chest CT scan model
│
├── Results/
│ ├── VGG_KUB_cancer_result.png # Kidney model training curves
│ ├── VGG_lung_cancer_result.png # Lung model training curves (unbalanced)
│ └── VGG_lung_cancer_balanced_results.png # Lung model training curves (balanced)
│
├── Webapp_Kidney_tumor/
│ ├── app.py # Flask application
│ ├── model.hdf5 # Deployed kidney model
│ ├── templates/index.html # Upload page
│ ├── templates/success.html # Results page
│ └── static/ # CSS and uploaded images
│
├── Webapp_lung_Cancer/
│ ├── app.py # Flask application
│ ├── model.hdf5 # Deployed lung cancer model
│ ├── templates/index.html # Upload page
│ ├── templates/success.html # Results page
│ └── static/ # CSS and uploaded images
│
├── kidney_tumor_dataset/ # Kidney CT scan dataset
├── lung_cancer_dataset/ # Lung CT scans (original)
├── lung_cancer_dataset_balanced/ # Lung CT scans (balanced)
│
├── requirements.txt
├── Interview_Guide.pdf # Detailed interview preparation guide
└── README.md

• Python 3.12+ (or 3.10 for original TF 2.10)
• 4GB+ RAM recommended for model loading

# Clone repository
git clone https://github.com/SriRammSS/tumor-classification.git
cd "tumor-classification"

# Create virtual environment
python3 -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate

# Install dependencies
pip install tensorflow tf-keras Flask==2.2.2 Pillow Werkzeug==2.2.3 Flask-Cors

cd Webapp_Kidney_tumor
python app.py
# → http://localhost:5000

cd Webapp_lung_Cancer
python app.py
# → http://localhost:5001

pip install jupyter notebook
jupyter notebook
# Open any .ipynb file to view/re-run training

This is a peer-reviewed, IEEE-published research project. The core IP is protected.

For academic collaboration, research reproduction, or dataset access, contact:

[email protected]

Include in your request:

• Your institution and role
• Intended use (research / teaching / benchmarking)
• Confirmation of non-commercial purpose

Commercial use requires a separate written licensing agreement.

If you use this work, you must cite the IEEE paper:

@inproceedings{sriramm2023crossorgan,
 author = {Sriramm, S. S. and Kamali, R. and Kishorkumar, S. M.
 and Venkatesh, K. V. Prasanna and Suguna, G.},
 title = {Cross Organ Bridge Transfer Learning for Lung Cancer Detection},
 booktitle = {2023 IEEE 11th Region 10 Humanitarian Technology Conference (R10-HTC)},
 year = {2023},
 pages = {876--883},
 doi = {10.1109/R10-HTC57504.2023.10461796}
}

Sri Ramm Sekar Sasirekha — Department of Electronics & Communication Engineering, Amrita School of Engineering, Coimbatore

Kamali R · S M Kishorkumar · K V Prasanna Venkatesh

Guide: Prof. Suguna G. — Dept. of ECE, Amrita School of Engineering