MOSInversion: Knowledge distillation-based incremental learning in organ segmentation using DeepInversion.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Current multi-organ segmentation (MOS) models struggle with incorporating new classes due to catastrophic forgetting.
• Existing incremental learning methods for MOS face challenges with memory requirements for 3D data and complex image synthesis.
• Catastrophic forgetting leads to performance degradation on previously learned classes in incremental learning settings.

Purpose of the Study:

• To develop an effective incremental learning framework for multi-organ segmentation (MOS) that addresses catastrophic forgetting.
• To enable MOS models to learn new organ classes progressively without significant performance loss on existing classes.
• To propose a method that is memory-efficient and suitable for 3D medical imaging data like CT scans.

Main Methods:

• Developed MOSInversion, an incremental learning framework utilizing diverse synthetic images generated from pre-trained models.
• Employed segmentation masks within MOSInversion to manipulate organ shape, location, and size for synthetic data generation.
• Evaluated the framework on three abdominal CT datasets: FLARE21, MSD, and KiTS19.

Main Results:

• The proposed MOSInversion framework successfully retained knowledge from previously learned data.
• Achieved state-of-the-art accuracy in multi-organ segmentation across the evaluated abdominal CT datasets.
• Demonstrated effective mitigation of catastrophic forgetting through the use of diverse synthetic images.

Conclusions:

• MOSInversion offers a robust solution for incremental learning in multi-organ segmentation, overcoming limitations of existing methods.
• The synthetic data generation approach effectively preserves performance on previously segmented organs.
• This framework shows significant promise for advancing AI in medical image analysis and clinical applications.