Project Introduction
The advent of AI has empowered image classification tasks such as melanoma detection with respect to skin lesions. However, the majority of such classifiers are trained on predominantly light skin tones. Techniques such as neural style transfer and self-supervised learning were implemented to create a novel dataset of dark skin toned skin lesion images that resulted in an improvement in classifier recall 0.623 to 0.787.
Lesion NST
Augment skin lesion datasets with generated dark skin toned images to improve melanoma classifier performance on darker skin tones
- Category: AI (Computer Vision)
- Client: Stanford AIMI
- Project date: 09/2022 - 12/2022
- Project Report: View
Contributions & Outcomes
I collected and pre-processed diverse skin lesion databases to create a consolidated dataset of cropped skin lesion images. The dataset includes corresponding ground truth labels stored in a .csv file. Additionally, I developed a TensorFlow pipeline for neural style transfer (NST) to generate dark skin tone lesion images. Subsequently, I designed a vanilla supervised classifier specifically for melanoma classification. As the dataset showed a higher proportion of benign images, I performed data augmentation techniques to enhance the accuracy of the melanoma classifier on both the original dataset and the NST-augmented dataset.
Technical Skills
- Computer Vision
- AWS Cloud Computing
- Python
- Open CV
- Tensorflow
- Research