Feature-conditioned neural network pre-training for skin cancer classification.
Abstract/Contents
- Abstract
- Despite tremendous advances in image recognition using convolutional neural networks over the past few years, the performance of deep learning based models on supervised classification is typically benchmarked using enormous labeled datasets comprising millions of images. Realistically, for many highly impactful classification tasks, such as medical image analysis, obtaining sufficient labeled data may be prohibitively expensive, time-consuming, or even impossible. However, it is often easier to obtain unlabeled image datasets, along with con-textual information that is related to the true desired labels. In this work, we propose two separate pre-training algorithms to leverage these datasets a GAN-based pre-training algorithm, and a purely CNN-based algorithm and evaluate their performance specifically for binary classification of skin lesions as either benign moles or cancerous melanomas. When tested against traditional model pre-training using the ImageNet dataset, both pre-training algorithms boost performance on the classification task, especially when smaller labeled datasets of only 5,000-10,000 images are available for training. Additionally, using the feature conditioned CNN-based pre-training algorithm, subsequent model performance exceeds both dermatologist and previous state-of-the-art results when evaluated on a test set of dermoscopic skin lesion images.
Description
Type of resource | text |
---|---|
Date created | 2017 |
Creators/Contributors
Author | Wong, Catherine |
---|---|
Advisor | Esteva, Andre |
Advisor | Thrun, Sebastian |
Subjects
Subject | Ben Wegbreit Prize for Best Undergraduate Honors Thesis |
---|---|
Subject | Artificial Intelligence |
Subject | Computer Science |
Genre | Thesis |
Bibliographic information
Access conditions
- Use and reproduction
- User agrees that, where applicable, content will not be used to identify or to otherwise infringe the privacy or confidentiality rights of individuals. Content distributed via the Stanford Digital Repository may be subject to additional license and use restrictions applied by the depositor.
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Preferred citation
- Preferred Citation
- Wong, Catherine. (2017). Feature-conditioned neural network pre-training for skin cancer classification. Stanford Digital Repository. Available at: https://purl.stanford.edu/nz411hx4576
Collection
Undergraduate Theses, School of Engineering
View other items in this collection in SearchWorksContact information
- Contact
- catwong@stanford.edu
Also listed in
Loading usage metrics...