Selected project
Demographic Bias in Medical Imaging.
A study of demographic bias in skin-cancer models trained on 10,015 dermoscopic images.
The project
What it does
This research project evaluates demographic bias in deep-learning models used for skin-cancer image classification. The work uses a dataset of 10,015 dermoscopic images and compares predictive behavior across demographic groups instead of treating aggregate accuracy as the only measure of model quality.
The study trains and evaluates convolutional neural-network architectures for lesion classification, then examines whether performance differences appear across the available demographic attributes. The goal is to surface fairness risks that can be hidden by a single overall performance score.
Core capabilities
What I built
- Studied model behavior across demographic groups.
- Worked with a dataset of 10,015 dermoscopic images.
- Used EfficientNet and ResNet50 in the modeling work.
Technology
Frameworks and tools
- Model architectures
- EfficientNet, ResNet50
- Domain
- Dermoscopic image classification, medical AI fairness
- Dataset
- 10,015 skin-lesion images
- Evaluation
- Model-performance and demographic-group comparisons
Process
How it works
- 01
Prepare and organize the dermoscopic image dataset for model training and evaluation.
- 02
Train transfer-learning image classifiers using established CNN architectures.
- 03
Measure classification performance across the full test set.
- 04
Break results down by demographic groups to identify disparities hidden by aggregate metrics.
- 05
Compare architectures and document the implications for fair medical-image modeling.