🔬 Cancer Classification using Convolutional Neural Networks

A deep learning project comparing three CNN architectures for multi-class cancer classification from histopathological images.

Overview

This project implements and compares three CNN approaches for classifying cancer types from medical images:

Model	Description	Test Accuracy
Baseline	Small CNN built from scratch, no regularization	81.4%
Enhanced	Deeper CNN with dropout, L2 regularization, and strided convolutions	92.9%
Transfer Learning	Frozen, ImageNet-pretrained VGG16 backbone + custom classification head	96.1%

See Results below for full per-class metrics, confusion matrices, and ROC curves.

Dataset

• Size: 25,000 histopathological images, resized to 120x120
• Classes: 5 (colon adenocarcinoma, colon benign tissue, lung adenocarcinoma, lung benign tissue, lung squamous cell carcinoma)
• Source: LC25000 Lung and Colon Histopathological Image Dataset (paper)

The dataset is too large to host in this repo. Download it from the link above and extract it locally (see Usage).

Project Structure

├── cancer_cnn/                # Core library
│   ├── config.py               # Paths, hyperparameters, label mapping
│   ├── data.py                 # Dataset loading, preprocessing, splitting
│   ├── visualize.py             # EDA plots (sample images, class distribution)
│   ├── evaluation.py            # Loss/accuracy curves, confusion matrices, ROC curves
│   └── models.py                # The three CNN architectures
├── scripts/
│   ├── prepare_dataset.py       # Raw images -> resized numpy arrays
│   └── train.py                  # Train one model end-to-end and save figures
├── results/figures/              # Saved output from each model
│   ├── eda/
│   ├── baseline/
│   ├── enhanced/
│   └── transfer_learning/
├── reports/
│   └── model_comparison_report.pdf   # Detailed write-up and analysis
└── requirements.txt

Usage

Install

pip install -r requirements.txt

1. Prepare the dataset

Download the LC25000 dataset from the link above, extract it, then convert it into resized numpy arrays:

python scripts/prepare_dataset.py \
    --raw-data-dir /path/to/lung_colon_image_set \
    --processed-dir data/processed

2. Train a model

python scripts/train.py --model baseline
python scripts/train.py --model enhanced
python scripts/train.py --model transfer

Each run trains the chosen architecture, prints test-set and validation-set classification reports, and saves loss/accuracy curves, a confusion matrix, and ROC curves to results/figures/<model>/.

Results

Exploratory Data Analysis

Sample images	Class distribution

Baseline Model — 81.4% test accuracy

Loss / Accuracy	Confusion Matrix	ROC Curve

Enhanced Model — 92.9% test accuracy

Adds dropout, L2 regularization, strided convolutions, and a tuned learning rate on top of the baseline architecture.

Loss / Accuracy	Confusion Matrix	ROC Curve

Transfer Learning Model (VGG16) — 96.1% test accuracy

A frozen, ImageNet-pretrained VGG16 backbone with the same custom classification head as the enhanced model.

Loss / Accuracy	Confusion Matrix	ROC Curve

Key Findings

• Transfer learning with a pretrained VGG16 backbone outperformed both custom architectures by a wide margin (96.1% vs. 81.4% baseline test accuracy), confirming that ImageNet features transfer well to histopathological images even though the two domains look nothing alike.
• Regularization (dropout + L2) and strided convolutions closed most of the gap between the baseline and transfer-learned model on their own, improving test accuracy from 81.4% to 92.9% without any pretrained weights.
• All three models struggle most on distinguishing lung adenocarcinoma from lung squamous cell carcinoma, the two visually closest classes in the dataset — this is where the confusion matrices show the most off-diagonal mass.

For the full write-up and analysis, see the detailed report.

Author

Ivaylo Papazov

License

This project is licensed under the MIT License.