MT-CooL: Multi-Task Cooperative Learning via Flat Minima Searching.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Multi-task learning (MTL) is established in natural image analysis but underexplored in medical imaging.
• Existing MTL methods often treat tasks as competing objectives, limiting performance.
• This research addresses the need for improved MTL strategies in the medical domain.

Purpose of the Study:

• To propose a novel multi-level optimization approach for MTL in medical imaging.
• To foster a cooperative learning environment where tasks benefit each other.
• To develop robust sub-models resilient to variations in other tasks.

Main Methods:

• Formulated MTL as a multi-level optimization problem, unlike traditional multi-objective formulations.
• Advocated for a cooperative strategy enabling tasks to leverage features from one another.
• Introduced a novel optimization strategy to find flat minima for robust sub-model learning.

Main Results:

• Demonstrated advantages on the OrganCMNIST dataset through parameter and comparison studies.
• Validated efficacy on three eye-related medical image datasets.
• Showcased superior performance compared to state-of-the-art MTL approaches.

Conclusions:

• The proposed multi-level optimization significantly enhances MTL performance in medical imaging.
• The cooperative approach leads to more robust and versatile multi-purpose models.
• This work opens new avenues for MTL applications in medical image analysis.