BreastNet: Early Breast Cancer Detection

BreastNet is a neural network based binary classifier designed to detect breast cancer at an early stage. Using nine numerical clinical features from the Breast Cancer Dataset (Kaggle: yasserh), it predicts whether a tumor is benign or malignant.

1. Problem Statement

Breast cancer is a major global health issue and early detection is critical for improving survival rates. Traditional diagnosis involves manual interpretation of cell measurements, which can be time consuming and sometimes inconsistent.

Goal: Develop a neural network that can automatically classify tumors as benign or malignant based on structured numeric data, offering a fast and reliable clinical decision support tool.

Why this matters:

• Early-stage detection significantly improves treatment outcomes.
• Algorithmic assistance can support pathologists and clinicians.
• Machine learning can analyze patterns too subtle for manual inspection.

---

2. Dataset Description

Source: Kaggle (Breast Cancer Dataset by yasserh) Link: https://www.kaggle.com/datasets/yasserh/breast-cancer-dataset

Dataset Details

• Total Samples: 699

• Total Features: 10

• Usable Features for Learning: 9

• Target Label Column: Class

  • 2 → Benign
  • 4 → Malignant

Feature List

These features describe visual and morphological characteristics of cell nuclei:

1. Clump Thickness
2. Uniformity of Cell Size
3. Uniformity of Cell Shape
4. Marginal Adhesion
5. Single Epithelial Cell Size
6. Bare Nuclei
7. Bland Chromatin
8. Normal Nucleoli
9. Mitoses

The Id column is removed because it does not contain predictive information.

Nature of Features

All nine features are integer scores between 1 and 10. Higher values generally correlate with higher malignancy risk.

---

3. Data Preprocessing Workflow

Your notebook follows a structured preprocessing pipeline:

Step 1: Load Dataset

Load CSV into pandas and inspect for missing values.

Step 2: Drop the Id Column

Id is irrelevant for prediction.

Step 3: Handle Missing Values

Commonly found in “Bare Nuclei”. Rows with missing values are removed or corrected.

Step 4: Encode the Class Label

Convert (2, 4) → (0, 1):

• 0 → Benign
• 1 → Malignant

Step 5: Train Test Split

Example: 80 percent training, 20 percent testing.

Step 6: Feature Scaling

Apply StandardScaler:

• Ensures all features are on similar scales
• Improves convergence during training

Step 7: Model Input Preparation

Convert scaled arrays into formats suitable for the neural network training loop.

---

4. Model Architecture (BreastNet)

BreastNet is a simple feed forward neural network built from scratch or using deep learning libraries.

Architecture Summary

Input Layer:

• 9 neurons (one per feature)

Hidden Layer 1:

• ~16 or 32 neurons
• Activation: ReLU

Hidden Layer 2 (optional):

• ~8 or 16 neurons
• Activation: ReLU

Output Layer:

• 1 neuron
• Activation: Sigmoid
• Outputs probability of malignancy

Why This Architecture?

• Dataset is small, so simple architecture avoids overfitting
• ReLU speeds up training
• Sigmoid enables true binary probability output
• Easy to interpret

Loss and Optimization

• Loss: Binary Cross Entropy
• Optimizer: Gradient Descent or Adam-like strategy
• Epochs: ~1300 (your notebook trained until convergence)
• Learning Rate: around 0.03

---

5. Training Behavior and Convergence

Your model shows a clear improvement in testing cost as training progresses:

Example Logs (taken from your notebook):

• Epoch 0:

  • Train Cost = 0.8021
  • Test Cost = 0.9719

• Epoch 300:

  • Train Cost = 0.1148
  • Test Cost = 0.1515

• Epoch 700:

  • Train Cost = 0.0614
  • Test Cost = 0.1170

• Epoch 990+:

  • Best test costs reach 0.1122

This demonstrates strong convergence and stable learning.

---

6. Final Evaluation Metrics (Extracted From Notebook)

Confusion Matrix

TP: 26  
TN: 83  
FP: 5  
FN: 0  

Direct Results

• Final Test Cost: 0.1122
• F1 Score: 0.9123

Derived Performance Metrics

Accuracy: (26 + 83) / (26 + 83 + 5 + 0) = 109 / 114 = 95.6 percent

Precision: 26 / (26 + 5) = 83.87 percent

Recall (Sensitivity): 26 / (26 + 0) = 100 percent

Specificity: 83 / (83 + 5) = 94.3 percent

F1 Score: = 91.23 percent

Interpretation

• Zero false negatives → the model never misses malignant cases
• High recall → clinically important for early cancer detection
• High specificity → good at distinguishing benign cases
• F1 score balances both precision and recall

This is a strong result given the small dataset size.

---

7. Key Observations

1. The model achieves excellent recall, critical for cancer screening.
2. Training stabilizes after about 900 epochs, with minimal test cost.
3. Misclassifications mainly occur as false positives, which is safer than false negatives.
4. Neural network captures relationships between features effectively.
5. Scaling features greatly helps training stability.

---

8. Project Conclusion

BreastNet demonstrates the effectiveness of neural networks in early breast cancer detection using structured numeric medical data.

The model:

• Achieves 95.6 percent accuracy
• Achieves 100 percent sensitivity to malignant cases
• Maintains strong generalization on test data
• Provides fast and consistent predictions

This project showcases how AI can augment medical decision making in real clinical settings.

---

9. Limitations

• Dataset is small and may not fully generalize to real-world populations.
• Based solely on tabular data; real diagnosis typically uses imaging (mammograms).
• The model is not medically validated.
• Potential overfitting due to small dataset and long training time.

---

10. Future Improvements

• Incorporate deeper networks or regularization techniques
• Add ROC AUC calculation and threshold tuning
• Use oversampling or class weights to handle imbalance
• Combine tabular data with image based models
• Deploy as a web-based prediction tool
• Validate on additional breast cancer datasets

---

11. Tools and Technologies Used

• Python
• pandas, NumPy
• StandardScaler
• Custom Neural Network
• Matplotlib for cost/metric visualization
• Jupyter Notebook
• Kaggle Breast Cancer Dataset

---