Learning to Segment Unseen Tasks In-Context
While deep learning models have become the predominant method for medical image segmentation, they are typically incapable of generalizing to new segmentation tasks---involving new anatomies, image modalities, or labels. For a new segmentation task, researchers will often have to prepare new task-sp...
Main Author: | |
---|---|
Other Authors: | |
Format: | Thesis |
Published: |
Massachusetts Institute of Technology
2024
|
Online Access: | https://hdl.handle.net/1721.1/156117 |
_version_ | 1811072875063410688 |
---|---|
author | Butoi, Victor Ion |
author2 | Guttag, John V. |
author_facet | Guttag, John V. Butoi, Victor Ion |
author_sort | Butoi, Victor Ion |
collection | MIT |
description | While deep learning models have become the predominant method for medical image segmentation, they are typically incapable of generalizing to new segmentation tasks---involving new anatomies, image modalities, or labels. For a new segmentation task, researchers will often have to prepare new task-specific models. This process is time-consuming and poses a substantial barrier for clinical researchers who often lack the resources and expertise to train neural networks.
We present UniverSeg, an in-context learning method for solving unseen medical segmentation tasks. Given a new image to segment, and a set of image-label pairs that define the task, UniverSeg can produce accurate segmentation predictions with no additional training. We demonstrate that UniverSeg substantially outperforms existing methods in solving unseen segmentation tasks, and thoroughly analyze important aspects of our proposed data, training, and inference paradigms. |
first_indexed | 2024-09-23T09:18:41Z |
format | Thesis |
id | mit-1721.1/156117 |
institution | Massachusetts Institute of Technology |
last_indexed | 2024-09-23T09:18:41Z |
publishDate | 2024 |
publisher | Massachusetts Institute of Technology |
record_format | dspace |
spelling | mit-1721.1/1561172024-08-15T03:52:43Z Learning to Segment Unseen Tasks In-Context Butoi, Victor Ion Guttag, John V. Dalca, Adrian V. Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science While deep learning models have become the predominant method for medical image segmentation, they are typically incapable of generalizing to new segmentation tasks---involving new anatomies, image modalities, or labels. For a new segmentation task, researchers will often have to prepare new task-specific models. This process is time-consuming and poses a substantial barrier for clinical researchers who often lack the resources and expertise to train neural networks. We present UniverSeg, an in-context learning method for solving unseen medical segmentation tasks. Given a new image to segment, and a set of image-label pairs that define the task, UniverSeg can produce accurate segmentation predictions with no additional training. We demonstrate that UniverSeg substantially outperforms existing methods in solving unseen segmentation tasks, and thoroughly analyze important aspects of our proposed data, training, and inference paradigms. S.M. 2024-08-14T19:52:44Z 2024-08-14T19:52:44Z 2024-05 2024-07-10T12:59:28.758Z Thesis https://hdl.handle.net/1721.1/156117 0000-0001-7118-1492 Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) Copyright retained by author(s) https://creativecommons.org/licenses/by-nc-nd/4.0/ application/pdf Massachusetts Institute of Technology |
spellingShingle | Butoi, Victor Ion Learning to Segment Unseen Tasks In-Context |
title | Learning to Segment Unseen Tasks In-Context |
title_full | Learning to Segment Unseen Tasks In-Context |
title_fullStr | Learning to Segment Unseen Tasks In-Context |
title_full_unstemmed | Learning to Segment Unseen Tasks In-Context |
title_short | Learning to Segment Unseen Tasks In-Context |
title_sort | learning to segment unseen tasks in context |
url | https://hdl.handle.net/1721.1/156117 |
work_keys_str_mv | AT butoivictorion learningtosegmentunseentasksincontext |