#DermaMNIST Classification using Simple CNN Classical 2D CNN baseline trained on binary DermaMNIST (Actinic Keratoses vs Basal Cell Carcinoma), serving as the classical counterpart in a QCNN vs CNN comparative study on skin lesion classification.

DermaMNIST Binary Classification using Classical CNN

A classical Convolutional Neural Network (CNN) implementation for binary skin lesion classification on the DermaMNIST dataset, distinguishing Actinic Keratoses and Intraepithelial Carcinoma (AKIEC) from Basal Cell Carcinoma (BCC).

This repository serves as the classical baseline in a comparative study against a Quantum Convolutional Neural Network (QCNN).

📄 Companion repository (QCNN): [Link to your QCNN repo]

---

📋 Project Overview

Property	Details
Task	Binary classification — AKIEC (Class 0) vs BCC (Class 1)
Dataset	DermaMNIST (MedMNIST v2)
Image size	28×28 pixels, grayscale
Framework	TensorFlow 2 / Keras
Total parameters	420,610
Test accuracy	~55%
Training	60 epochs, batch size 64
Platform	Google Colab

---

🗂️ Dataset

DermaMNIST is part of the MedMNIST v2 benchmark collection, derived from the HAM10000 dermoscopy dataset (Tschandl et al., 2018). Images are 28×28 pixels in RGB format, standardised across 7 lesion categories.

This project uses a binary subset of two classes:

Label	Class	Description
0	AKIEC	Actinic Keratoses & Intraepithelial Carcinoma — precancerous lesion
1	BCC	Basal Cell Carcinoma — most common skin cancer

The dataset is accessed via the medmnist Python package, which handles downloading, caching, and loading automatically.

---

🏗️ Model Architecture

Input (28×28×1) │ ▼ Conv2D (32 filters, 3×3, same padding) → ReLU → MaxPooling2D (2×2) │ ▼ Conv2D (64 filters, 3×3, same padding) → ReLU → MaxPooling2D (2×2) │ ▼ Flatten (3136 units) │ ▼ Dense (128 units, ReLU) → Dropout (0.3) │ ▼ Dense (2 units, Softmax) → [P(AKIEC), P(BCC)]

Layer-wise parameter count

Layer	Output Shape	Parameters
Conv2D (32 filters)	(28, 28, 32)	320
MaxPooling2D	(14, 14, 32)	0
Conv2D (64 filters)	(14, 14, 64)	18,496
MaxPooling2D	(7, 7, 64)	0
Flatten	(3136,)	0
Dense (128)	(128,)	401,536
Dropout (0.3)	(128,)	0
Dense (2, softmax)	(2,)	258
Total		420,610

---

⚙️ Training Configuration

Parameter	Value
Loss function	Categorical cross-entropy
Optimiser	Adam (lr=0.001, weight_decay=0.0001)
Epochs	60
Batch size	64
Train/Val split	80% / 20%
Class imbalance	sklearn compute_class_weight('balanced')
Gradient method	Backpropagation

Data Augmentation

Applied online during training only via Keras ImageDataGenerator:

Augmentation	Setting
Rotation	±10°
Horizontal flip	Enabled (50% probability)
Zoom	±10%

---

📊 Results

Metric	Value
Test Accuracy	~55%
Chance baseline	50%
Above chance	+5 percentage points

The model correctly learns discriminative features from DermaMNIST dermoscopy images, exceeding the 50% chance baseline. The modest accuracy reflects the inherent difficulty of distinguishing AKIEC from BCC on 28×28 pixel grayscale images with a small training set (~670 samples).

---

🔁 How to Run

Option 1 — Google Colab (Recommended)

Click the badge below to open directly in Colab:

Option 2 — Run Locally

Step 1 — Clone the repository

git clone https://github.com/your-username/your-repo-name.git
cd your-repo-name

Step 2 — Install dependencies

pip install tensorflow medmnist scikit-learn numpy matplotlib

Step 3 — Run the notebook

jupyter notebook MSc_simple_CNN_dermaMNIST.ipynb

---

📦 Requirements

tensorflow>=2.0 keras medmnist scikit-learn numpy matplotlib

📁 Repository Structure

├── MSc_simple_CNN_dermaMNIST.ipynb # Main Colab notebook ├── README.md # This file

---

🔬 Preprocessing Pipeline

Raw DermaMNIST images (uint8, [0, 255]) │ ▼ Load as grayscale (as_rgb=False) → shape: (N, 28, 28, 1) │ ▼ Normalise to [0.0, 1.0] → divide by 255.0 │ ▼ Filter for AKIEC (0) and BCC (1) only │ ▼ One-hot encode labels → [1,0] for AKIEC, [0,1] for BCC │ ▼ 80/20 train/validation split │ ▼ Apply class-weight balancing │ ▼ Online augmentation via ImageDataGenerator (training only) │ ▼ Feed to CNN → train for 60 epochs

---

📈 Comparison with QCNN

This repository is the classical baseline in a QCNN vs CNN comparative study:

Property	CNN (This repo)	QCNN (Companion repo)
Framework	TensorFlow / Keras	PennyLane + PyTorch
Input	Raw 28×28×1 image	16-dim autoencoder latent vector
Trainable parameters	420,610	51
Test accuracy	~55%	~60%
Gradient method	Backpropagation	Parameter-shift rule
Training duration	60 epochs	200 steps
Hardware	GPU (Colab)	CPU (Colab)

The QCNN achieves ~5 percentage points higher accuracy with 8,247× fewer parameters, demonstrating competitive quantum representational capacity in a low-data medical imaging regime.

---