Gastric Cancer Histopathology Image Classification

This project builds a Convolutional Neural Network (CNN) using TensorFlow and Keras to classify histopathological tissue images into different gastric tissue types. The trained model is deployed using Streamlit on Hugging Face Spaces for interactive prediction.

---

📂 Dataset

Name: Gastric Cancer Histopathology Tissue Image Dataset
Source: Kaggle Dataset

Folder Structure (after extraction):

gastric_data/
├── HMU-GC-HE-30K/
│   ├── all_image/
│   │   ├── ADI/
│   │   ├── DEB/
│   │   ├── LYM/
│   │   ├── MUC/
│   │   ├── MUS/
│   │   ├── NOR/
│   │   ├── STR/
│   │   └── TUM/

---

📌 Project Goals

• Classify different gastric tissue types from histopathology images
• Handle multiple classes (8 classes)
• Build and train a CNN model with early stopping and checkpointing
• Deploy for real-time inference using Streamlit on Hugging Face

---

🧪 Steps Performed

1. Dataset Exploration & Visualization

• Examined the folder and class distribution.
• Displayed sample images from each tissue class.
• Checked for class imbalance.

2. Data Preprocessing

• Resized images to 224x224 pixels.
• Normalized pixel values to [0, 1].
• Created training and validation datasets using image_dataset_from_directory().

3. Model Architecture

model = Sequential([
    Conv2D(32, (3,3), activation='relu', input_shape=(224,224,3)),
    MaxPooling2D(2,2),
    Conv2D(64, (3,3), activation='relu'),
    MaxPooling2D(2,2),
    Conv2D(128, (3,3), activation='relu'),
    MaxPooling2D(2,2),
    Flatten(),
    Dropout(0.5),
    Dense(128, activation='relu'),
    Dense(8, activation='softmax')  # 8-class classification
])

4. Model Compilation

model.compile(optimizer='adam',
              loss='categorical_crossentropy',
              metrics=['accuracy'])

5. Training Setup

• Trained for up to 20 epochs with EarlyStopping.
• Used ModelCheckpoint to save the best model (bestmodel.keras).

6. Performance Evaluation

• Calculated accuracy, precision, recall, F1-score.
• Generated a confusion matrix.
• Identified and analyzed misclassified images.

---

📊 Results

Metric	Value (Approx.)
Training Accuracy	↑ steadily up to ~0.85
Validation Accuracy	Stabilized ~0.70–0.75
Validation Loss	Controlled well with early stopping

The model showed good generalization. Some confusion between certain tissue types remains, indicating scope for further fine-tuning.

---

🧠 Key Learnings

• CNNs can effectively distinguish different tissue types in medical images.
• EarlyStopping and ModelCheckpoint improve model robustness.
• Deploying on Hugging Face makes the project easily accessible.

---

🖥️ Deployment

The trained model is deployed using Streamlit on Hugging Face Spaces.

🔗 Access the deployed app here:
👉 Gastric Cancer Detection App

Deployment Steps:

1. Trained model saved as bestmodel.keras.
2. Streamlit app (app.py) created to upload and predict images.
3. requirements.txt included for dependency management.
4. Uploaded all files (app.py, bestmodel.keras, requirements.txt) to Hugging Face Spaces.
5. App runs instantly in the browser!

---

🧰 Dependencies

• Python 3.8+
• TensorFlow
• Streamlit
• NumPy
• Pillow
• scikit-learn

---

⚠️ Ethical Considerations

• AI-based diagnostic tools should assist, not replace medical professionals.
• Clinical validation is mandatory before real-world deployment.
• Biases and errors in AI models must be handled responsibly.

---

📃 License

This project is intended for educational and research purposes only.

---

✍️ Author

Shubham Boral
CSE Pre-Final Year Student | Passionate about AI, Medical Imaging, and Cybersecurity