Building medical image classifiers with very limited data using segmentation networks.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Deep learning in medical imaging is limited by small annotated datasets.
• ImageNet pre-training offers benefits but has limitations in computational cost and performance due to model complexity and image size constraints.
• Shared morphological features across organ classification tasks present an opportunity for improved performance with limited samples.

Purpose of the Study:

• To propose a novel framework for building medical image classifiers using features extracted from segmentation networks.
• To leverage curriculum learning principles by first learning shape and structural concepts before complex classification.
• To enhance classification performance in medical imaging tasks with limited annotated data.

Main Methods:

• A strategy for medical image classification using features from pre-trained segmentation networks was developed.
• The approach utilizes curriculum learning by training on simpler segmentation tasks before complex classification.
• The framework was evaluated on 3D brain tumor classification and 2D cardiac semantic level classification problems.

Main Results:

• Achieved 82% accuracy on a 3D brain tumor classification task (91 training, 191 testing samples).
• Achieved 86% accuracy on a 2D cardiac semantic level classification task (108 training, 263 testing samples).
• Demonstrated superior performance compared to ImageNet pre-trained classifiers and models trained from scratch.

Conclusions:

• The proposed framework effectively utilizes segmentation network features for improved medical image classification with limited data.
• This curriculum learning-inspired approach offers a viable solution to dataset size limitations in medical deep learning.
• The method shows significant potential for enhancing diagnostic accuracy in various medical imaging applications.