Generalizable deep learning framework for 3D medical image segmentation using limited training data.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence in Medicine

Background:

• Medical image segmentation is crucial for clinical applications like 3D printing and surgical planning.
• Manual segmentation is time-consuming and prone to variability.
• Deep learning offers automation but typically requires large datasets and significant GPU power.

Purpose of the Study:

• To introduce a robust deep learning framework for 3D medical segmentation.
• To enable high-performance segmentation with limited data and computational resources.
• To demonstrate applicability across diverse clinical scenarios.

Main Methods:

• Development of a novel deep learning framework for 3D medical image segmentation.
• Training and validation on a small number of subjects across various anatomical structures.
• Utilizing a minimal set of hyper-parameters and augmentation settings.

Main Results:

• Achieved an average Dice score of 92% (SD = ±0.06) across diverse organs and tissues.
• Demonstrated high performance in six distinct clinical applications (orthopedics, orbital, mandible CT, cardiac CT, fetal MRI, lung CT).
• Framework proved effective even with limited training data and computational resources.

Conclusions:

• The proposed deep learning framework offers a resource-efficient solution for 3D medical segmentation.
• It overcomes limitations of traditional deep learning methods, enhancing accessibility in healthcare.
• The approach facilitates advanced visualization, surgical planning, and 3D printing in clinical practice.