Back to Projects

Medical AI / Deep Learning· 2024

Skin Cancer Classification

VGG-16 transfer learning vs custom CNN for melanoma detection

87%

VGG-16 Accuracy

85

Dataset Size

4

Augmentation Types

2

Models Compared

Problem

Early detection of skin cancer (melanoma) is critical for patient survival. Dermatologists face high volumes of skin lesion examinations, and misdiagnosis can have severe consequences.

Solution

Implemented and compared two deep learning approaches — VGG-16 transfer learning and a custom CNN — to classify skin lesions as benign or malignant. Demonstrated the effectiveness of transfer learning on limited medical imaging data.

Tech Stack

TensorFlowKerasVGG-16NumPyPandasMatplotlibPIL

Key Features

VGG-16 transfer learning with frozen ImageNet weights

Custom 3-layer CNN architecture for comparison

Data augmentation (rotation, zoom, shear, flip) for small dataset handling

Results

VGG-16 transfer learning outperformed custom CNN on limited data

Successful binary classification of benign vs malignant lesions

Demonstrated transfer learning advantage on small medical datasets

Visualizations

Dataset Samples — Benign and malignant skin lesion examples from ISIC

Dataset Samples — Benign and malignant skin lesion examples from ISIC

Age Distribution by Anatomical Site — Violin plot of patient demographics

Age Distribution by Anatomical Site — Violin plot of patient demographics

Dataset Overview

Class Distribution — 50 benign vs 35 malignant images

Class Distribution — 50 benign vs 35 malignant images

Class Ratio — 58.8% benign, 41.2% malignant

Class Ratio — 58.8% benign, 41.2% malignant