Synchronous Medical Image Augmentation framework for deep learning-based image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning (DL) models are crucial for medical image analysis, but their performance relies heavily on large, diverse datasets.
• Medical image acquisition faces challenges like data scarcity, class imbalance, and high annotation costs.
• Existing image augmentation methods often focus on classification, neglecting the specific needs of image segmentation tasks.

Purpose of the Study:

• To develop a novel framework for medical image augmentation specifically tailored for deep learning-based image segmentation.
• To ensure synchronization between augmented image samples and their corresponding labels.
• To address data scarcity and imbalance issues in medical imaging datasets for segmentation.

Main Methods:

• Proposed a Synchronous Medical Image Augmentation (SMIA) framework with two modules: stochastic transformation and synthesis.
• The transform-based module generates augmented images and paired tissue segments simultaneously using selected SMIA factors.
• The synthesis-based module creates new medical images by replacing tissues with augmented counterparts, preserving medical context.

Main Results:

• The SMIA framework successfully generated category-balanced and diverse training data.
• Experiments on bone marrow smear and dermoscopic images demonstrated the framework's effectiveness.
• The proposed augmentation method positively impacted the performance of DL-based image segmentation models.

Conclusions:

• The SMIA framework effectively enhances medical image datasets for DL segmentation.
• Synchronized augmentation of images and labels is crucial for improving segmentation model accuracy.
• This approach offers a viable solution to data limitations in medical image segmentation.