Year: 2025
Type: dissertation
Abstract: Automatic polyp segmentation using deep learning models has great potential to reduce missed polyp detections and improve colorectal cancer outcomes.However, these models often pick up patterns from the training data that are unrelated to the polyps themselves.For example, they might focus on specific annotation styles introduced by the imaging device or on demographic and anatomical variations across different populations.These factors can influence the model predictions without providing any meaningful information about the lesions, thus making them biased and leading to poor performance on new data.Consequently, such models struggle when faced with distribution shifts, which hinders their proper integration into clinical settings and limits their potential to improve diagnostic accuracy.This dissertation explores methods to address these biases, aiming to enhance the reliability of deep learning models across diverse data distributions and to promote their effective use in clinical applications.The primary goal is to enhance polyp segmentation by increasing feature robustness, encouraging models to focus on meaningful polyp-related features, while reducing reliance on irrelevant cues.This is explored through the application of color invariance mechanisms and self-supervised pretraining, which has been shown to produce more meaningful feature representations when compared to label-based methods.The robustness of these features is further enhanced through the self-attention mechanisms used in Vision Transformers.By integrating these methods, we propose a simple yet effective architecture that leverages high-quality learned features to improve current state-of-the-art generalization performance.In sum, the contributions of this thesis aim to improve the performance of polyp segmentation models across varying data distributions, bringing them closer to practical integration in clinical settings.
Author: Carla Sofia Barros Monteiro
Institution: Universidade do Porto
Language: English
Subfield: Oncology
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7111791409
