Cancer Cell Classification Using Deep Learning

This project aims to classify cancerous and non-cancerous cell images using deep learning models. The primary objective was to train and evaluate multiple CNN architectures to determine the most effective model for accurate classification.

Project Overview

This repository contains:

• Exploratory Data Analysis (EDA) to understand dataset distribution and augmentation effects.
• Baseline CNN Model as a reference classifier.
• Transfer Learning Models (EfficientNet, MobileNet, and ResNet) to improve classification performance.
• Evaluation Metrics & Comparisons including confusion matrices, precision-recall reports, and ROC curves.

Steps for Execution

Drive Link: https://drive.google.com/drive/folders/1FWv_40wRqUOBDmpoIA8TR1UK8EsO29yV?usp=sharing

Step 1: Exploratory Data Analysis (EDA)

EDA provides insights into dataset distribution and augmentation effects. To replicate EDA, refer to:

• EDA Report: EDA.md
• Jupyter Notebook: src/CancerCellClassification_EDA.ipynb
• HTML Report: src/CancerCellClassification_EDA.html

Step 2: Train & Evaluate Models

Model training and evaluation are documented in:

• Training & Validation Notebook: src/Train_Test_Validate_Phase1.ipynb
• HTML Report: Tsrc/rain_Test_Validate_Phase1.html

To replicate results, run the Jupyter notebooks in the order provided.

Dataset Sample Visualizations

Training Data Samples

Validation Data Samples

Test Data Samples

Model Training & Evaluation

Baseline Model

---

The baseline CNN was trained as a reference.

• Training Progress

  

• Classification Report

  

• Confusion Matrix

  

• Sample Predictions

  

Transfer Learning Models

I fine-tuned EfficientNetB0, MobileNetV2, and ResNet50 for performance improvement.

EfficientNet

---

• Training Progress

  

• Classification Report

  

• Confusion Matrix

  

• Sample Predictions

  

MobileNet

---

• Training Progress

  

• Classification Report

  

• Confusion Matrix

  

• Sample Predictions

  

ResNet

---

• Training Progress

  

• Classification Report

  

• Confusion Matrix

  

• Sample Predictions

  

Model Comparison

---

To compare the models, I analyzed ROC curves and classification metrics.

• ROC Curve

  

Conclusion

---

• EfficientNetB0 provided the best balance betIen accuracy, recall, and computational efficiency.
• MobileNetV2 provided competitive performance with loIr resource requirements.
• ResNet50 provided high precision but loIr recall, indicating a tendency to misclassify malignant cases.

This project demonstrates how deep learning can be effectively utilized in medical image classification. The structured pipeline ensures reproducibility and can be extended to other classification tasks.