Self-supervised learning for medical image data with anatomy-oriented imaging planes.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Self-supervised learning (SSL) is vital for pretraining deep networks on unlabeled data.
• Effective transfer learning relies on the relevance between pretraining and target tasks.
• Existing SSL methods often overlook anatomy-oriented imaging planes in medical data.

Purpose of the Study:

• To develop novel pretext tasks for SSL in medical imaging, specifically for anatomy-oriented views.
• To leverage the spatial relationships of imaging planes for improved representation learning.
• To enhance deep network performance in medical image analysis through specialized pretraining.

Main Methods:

• Proposed two complementary pretext tasks based on spatial relationships of imaging planes.
• Task 1: Regressing intersecting lines to learn relative orientation between planes.
• Task 2: Regressing relative slice locations within parallel planes for positional understanding.

Main Results:

• Demonstrated effectiveness of proposed pretext tasks on cardiac and knee imaging datasets.
• Achieved significantly improved performance on semantic segmentation and classification tasks.
• Outperformed other recent SSL approaches in transfer learning scenarios.

Conclusions:

• The proposed pretext tasks are effective for pretraining deep networks on anatomy-oriented medical images.
• These methods offer a straightforward yet powerful approach to enhance medical image analysis.
• The combined multitask learning strategy shows promise for superior representation learning.