CernovaAI/CANet-v1.4

---

Multi-Cancer Lymphoma Classification with Convolutional Neural Networks (CNN)

📌 Overview

This repository contains an end-to-end deep learning pipeline developed in Python using TensorFlow and Keras for the automated classification of lymphoma subtypes within a multi-cancer dataset. The project leverages Convolutional Neural Networks (CNNs) to perform supervised image classification on histopathological cancer images, aiming to provide a robust and scalable solution for medical imaging analysis.

The pipeline encompasses:

• Data ingestion and preprocessing with ImageDataGenerator
• Training/validation split and augmentation
• Definition and compilation of a deep CNN architecture
• Training with real-time performance evaluation
• Model persistence (.h5 file format) for later inference
• Custom prediction utility with visualization

This repository is intended for medical AI researchers, machine learning engineers, and healthcare data scientists who seek to apply convolutional neural networks for diagnostic support in oncology.

---

📂 Dataset Information

The dataset used in this project is located at:

/kaggle/input/multi-cancer/Multi Cancer/Multi Cancer/Lymphoma

This directory contains subfolders representing different classes of lymphoma and potentially other cancer subtypes. The directory structure is expected to be of the form:

Lymphoma/
    ├── Class_A/
    │   ├── img_1.jpg
    │   ├── img_2.jpg
    │   └── ...
    ├── Class_B/
    │   ├── img_3.jpg
    │   └── ...
    └── Class_C/
        ├── img_4.jpg
        └── ...

• Each subfolder corresponds to one diagnostic class.
• The model automatically infers class labels from these subdirectories.

---

⚙️ Dependencies

This project requires the following core dependencies:

• Python 3.8+
• TensorFlow 2.x
• Keras (integrated with TensorFlow)
• NumPy
• Matplotlib

To install dependencies:

pip install tensorflow numpy matplotlib

If running on Kaggle or Google Colab, these libraries are already pre-installed.

---

🧩 Code Structure

The main script (train.py or notebook cell) is divided into logical sections:

1. Imports

   • Standard libraries (os, numpy)
   • Scientific libraries (matplotlib)
   • Deep learning libraries (tensorflow, keras, layers)

2. Data Pipeline

   • Data preprocessing with ImageDataGenerator
   • Automatic normalization of pixel intensities (rescale=1./255)
   • Splitting into training (90%) and validation (10%)

3. Model Architecture

   • A sequential CNN architecture with the following layers:

     • Conv2D (32 filters, 3×3 kernel, ReLU)
     • MaxPooling2D (2×2)
     • Conv2D (64 filters, ReLU)
     • MaxPooling2D (2×2)
     • Conv2D (128 filters, ReLU)
     • MaxPooling2D (2×2)
     • Flatten
     • Dense (512 units, ReLU)
     • Dense (softmax output for multi-class classification)

4. Compilation

   • Optimizer: Adam
   • Loss Function: Categorical Crossentropy
   • Metrics: Accuracy

5. Training

   • Training via model.fit()
   • epochs=10
   • Validation data monitoring

6. Model Persistence

   • Final trained model is saved as model5.h5

7. Prediction Utility (guess() function)

   • Takes an input image path
   • Resizes and normalizes the image
   • Performs forward propagation using the trained model
   • Outputs the predicted class with corresponding visualization

---

🔬 Methodology

The approach relies on supervised learning using CNNs for image recognition.

• Feature Extraction: Convolutional and pooling layers learn hierarchical spatial representations of cancerous tissue patterns.
• Classification: Dense layers map these features into probabilistic class predictions.
• Normalization: All images are rescaled to [0,1] for stable gradient descent.
• Generalization: Validation set (10%) monitors overfitting and ensures out-of-sample reliability.

This is a baseline model, and can be extended with:

• Data Augmentation (rotation, zoom, shear, flips)
• Transfer Learning (e.g., VGG16, ResNet50, EfficientNet)
• Regularization (Dropout, L2 penalty)
• Hyperparameter Optimization (learning rate, batch size tuning)

---

📊 Training Performance

• Epochs: 10
• Batch Size: 32
• Image Size: 150×150 (RGB channels)
• Optimizer: Adam (adaptive learning rate)
• Loss Function: Categorical Crossentropy
• Evaluation Metric: Accuracy

Performance metrics will be printed during runtime and can be plotted for visualization. Example outputs include training/validation accuracy and loss curves.

---

🧪 Inference Example

Using the custom guess() function:

from tensorflow.keras.models import load_model

# Load model
model = load_model("model5.h5")

# Predict on new image
guess("example_image.jpg", model, train_generator.class_indices)

Expected Output:

• The image is displayed.
• The title above the image indicates the predicted lymphoma subtype.

---

📌 Applications

• Medical Decision Support: Assisting oncologists in rapid and preliminary diagnosis of lymphoma subtypes.
• Research: Benchmarking CNN performance on histopathological datasets.
• Education: Teaching medical students and engineers about AI applications in pathology.

⚠️ Disclaimer: This model is for research and educational purposes only. It is not a substitute for professional medical diagnosis. Clinical deployment requires extensive validation, regulatory approval, and rigorous testing.

---

🚀 Future Improvements

1. Integrating transfer learning for improved accuracy.
2. Expanding dataset size and diversity.
3. Hyperparameter optimization with automated search tools.
4. Deploying as a web application (e.g., Flask, FastAPI, Streamlit).
5. Exporting to TensorFlow Lite or ONNX for mobile/edge deployment.

---

🏆 Conclusion

This project demonstrates the development of a robust, reproducible, and interpretable CNN-based classification model for multi-cancer (lymphoma) image analysis. It provides a solid foundation for further advancements in AI-driven oncology research.

By following the modular design of this repository, researchers can:

• Reproduce experiments
• Extend the architecture
• Adapt the pipeline for other cancer datasets

This repository bridges the gap between machine learning engineering and medical research, contributing towards a future where AI supports healthcare professionals in delivering faster, more accurate, and more reliable diagnoses.

---