This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
MOST: MR reconstruction Optimization for multiple downStream Tasks via continual learning
Created by
Haebom
Author
Hwihun Jeong, Se Young Chun, Jongho Lee
Outline
This paper points out that while deep learning-based MRI reconstruction methods focus on high-quality image generation, the impact on subsequent tasks (e.g., segmentation) that utilize the reconstructed images is often overlooked. The approach of linking separately trained reconstruction networks with subsequent task networks has been shown to result in poor performance due to error propagation and domain gaps between training datasets. To alleviate this problem, a subsequent task-oriented reconstruction optimization for a single subsequent task is proposed. It is not trivial to extend this optimization to a multi-task scenario. In this paper, we extend this optimization to multiple sequentially introduced subsequent tasks and show that a single MR reconstruction network can be optimized for multiple subsequent tasks via continuous learning (MOST). MOST overcomes catastrophic forgetting by integrating replay-based continuous learning and image-guided loss techniques. Comparative experiments demonstrate that MOST outperforms reconstruction networks without fine-tuning, reconstruction networks with simple fine-tuning, and existing continuous learning methods. The source code is available at https://github.com/SNU-LIST/MOST .
We contribute to improving the performance of subsequent tasks by presenting a continuous learning method (MOST) that optimizes a single MR reconstruction network for multiple subsequent tasks.
◦
Combining replay-based continuous learning and image-guided loss to solve the critical forgetting problem.
◦
We present the possibility of efficiently training a single network for various follow-up tasks.
•
Limitations:
◦
There may be limitations in the applicability of simultaneous multi-task learning, as it adds tasks sequentially.
◦
As a performance evaluation for specific medical image data, generalization performance to other types of data or tasks requires further study.
◦
Lack of analysis of the complexity and computational cost of the proposed MOST algorithm.
