Interactive Image Selection and Training for Brain Tumor Segmentation Network.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Deep learning for medical imaging

Background:

• Medical image segmentation is crucial but requires extensive annotated data for deep learning models.
• Diverse pathologies like brain tumors present challenges due to variations in size and shape.
• Traditional deep learning methods demand significant computational resources and large datasets.

Purpose of the Study:

• To develop an interactive, expert-guided image selection and training method for medical image segmentation.
• To leverage the Feature Learning from Image Markers (FLIM) methodology for efficient model training.
• To reduce the dependency on large, fully annotated datasets in medical image segmentation tasks.

Main Methods:

• Implementation of an interactive system for selecting training images based on expert input.
• Application of the Feature Learning from Image Markers (FLIM) approach for training convolutional layers without backpropagation.
• Training the encoder of a U-shaped network using a small, interactively selected set of images.

Main Results:

• The proposed FLIM-based interactive method achieved performance comparable to manual image selection.
• The method successfully trained a U-shaped network encoder with a minimal dataset.
• The trained model surpassed the performance of a U-shaped network trained with backpropagation using all available images.

Conclusions:

• Interactive image selection combined with FLIM offers an efficient alternative for training medical image segmentation models.
• This methodology significantly reduces the need for large annotated datasets, making deep learning more accessible for diverse medical imaging applications.
• The approach demonstrates the potential of incorporating expert knowledge into the deep learning pipeline for improved efficiency and performance.