Year: 2016
Type: dissertation
Abstract: This project presents an exploratory study on the potential of using convolutional neural networks (CNNs) to aid in polyp detection tasks. Early polyp detection is crucial for patient survival and it could be very useful for clinicians to have a tool that could aid clinicians in real time to identify polyps in the image. The main challenge we faced was to work in a domain where publicly available datasets are scarce and, moreover, they are not balanced, we aim to assess the potential of using pre-trained networks and fine-tuning to take profit of the capa-bilities of CNNs. We propose two different strategies: classification using the whole image and patch-based classification. As the databases that we use to validate our methodology are clearly not balanced, we also explore the use of early image augmentation techniques to increase dataset size. Preliminary results indicate the potential of our proposal and assess quantitatively the impact of each of the modifications we propose to the baseline. More precisely, the use of patch-based classification outperforms image-based one, and there is a proven impact of supressing image borders and performing data augmentation. Finally, our results shows the importance of having balanced datasets for both training and testing of the network.
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2529351194