Seminario de Informática ‘Mirian Andrés’. Towards Improving Medical Image Semantic Segmentation using Model Soups

Semantic segmentation is a computer vision task that involves assigning each pixel in an image to a predefined category or class, and nowadays is mainly approached by developing deep learning models. These models play a crucial role in medical imaging; however, given the critical nature of medical applications, it is essential that each pixel of an image is correctly classified. In order to tackle this problem, several techniques are commonly used but they often entail high computational costs and require the generation of intermediate models that are usually discarded if they do not obtain the best results.

Model soups take advantage of these intermediate models to improve their performance. This approach aims to maximize metrics by averaging the weights of multiple models trained with different hyperparameters. In this work, we have developed a library to facilitate the creation of model soups for semantic segmentation independently of the underlying architecture of the segmentation models. We finally evaluate whether medical semantic segmentation performance can be improved using model soups built from intermediate checkpoints and different weight-averaging strategies.