Final Degree Project

Study and Improvement of Image Segmentation Techniques through Self-Supervised Learning

Author: Ismael Tse Perdomo Rodríguez Mentors: Félix Fuentes y Guillermo Iglesias

Asbtract

Despite the remarkable success in natural language processing and image classification, the self-supervised paradigm has not been particularly impactful in the domain of semantic segmentation. This Final Degree Project aims to study and enhance current techniques for laparoscopic image segmentation using self-supervised learning, with the objective of improving clinical outcomes and patient safety.

A self-supervised pre-training pipeline was developed, integrating both discriminative and generative approaches to extract essential information from images without the need for manual labeling. This method was rigorously validated through extensive experiments, and its performance was compared to two supervised baselines representing the main traditional architectures.

The results demonstrate a significant enhancement in laparoscopic image segmentation, with a substantial reduction in processing time by up to 40% compared to supervised techniques. These improvements are especially notable in complex clinical scenarios where image variability is high and the margin for error is minimal, achieving a \textit{Dice Coefficient} of 0.95.

This methodology shows great promise for application in real clinical settings, paving the way for continued advancements in the precision and efficiency of image-assisted surgical procedures.

Citations

@mastersthesis{citekey,
  title = {Estudio y Mejora de Técnicas de Segmentación en Imágenes Laparoscópicas
  a través del Aprendizaje Auto Supervisado},
  author = {Ismael Tse Perdomo y Felix Fuentes y Guillermo Iglesias},
  school = {E.T.S. de Ingeniería de Sistemas Informáticos},
  year = {2024},
  month = {7},
  type = {Proyecto Fin de Grado}
}