Few-shot Learning in Medical Image Segmentation
Keywords
No Thumbnail Available
Authors
Issue Date
2024-03-19
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
There are various ways to train artificial intelligence models, and the most effective ones are those that
require a vast amount of time and data. A more efficient approach to machine learning has recently come
to light, which is called few-shot learning. What this approach attempts is to perform at a level close to
state-of-the-art techniques, but with far less data to train on. In this research, the goal was to determine
whether few-shot learning could be used with medical images, and computed tomography (CT) scans were
used to test this. A data set focusing on the thorax and abdomen was created by combining 3 individual
data sets. A base model using a nnU-Net architecture was trained on most of the 2D data, and subsequent
few-shot models were trained on small portions of the data set specifically focusing on either a specific
organ, muscle, or bone, which the base model had not trained on. Multiple training variations have been
tested, which were adjustments to the learning rate, loss function, or data set. It has been found that
the usage of a base model, as well as an increase in training data, resulted in a positive effect on the
performance. Moreover, it has been found that only training on data that incorporates the relevant class,
or simply increasing the amount of training data, highly impacts the convergence speed of the model during
training. For the training variations, lowering the learning rate barely impacted the performance of the
models, whereas freezing the learning rate except for the last layer of the model negatively impacted the
performance of the models. Both adjusting the loss function and removing training images without the
relevant class resulted in a slightly worse performance. Although the results themselves were weak, they
do provide information on what could be done to improve subsequent training using a 2D nnU-Net model.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen