Semi Supervised Learning with Deep Embedded Clustering for Image Classification and Segmentation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) typically require large labeled datasets for accurate model training.
• Acquiring labeled data, especially for specialized tasks like medical image segmentation, is often expensive and time-consuming.
• Semi-supervised learning (SSL) methods offer a solution by utilizing both limited labeled data and abundant unlabeled data.

Purpose of the Study:

• To introduce a flexible semi-supervised learning framework, Semi-Supervised Learning based on Deep Embedded Clustering (SSLDEC).
• To combine deep convolutional neural networks for feature representation with deep embedded clustering for data assignment.
• To reduce the dependency on large labeled datasets in DNNs for tasks like image classification and segmentation.

Main Methods:

• SSLDEC iteratively learns feature representations by alternating between labeled and unlabeled data.
• It computes target distributions from predictions, using labeled samples to maintain consistency and tune the model.
• The framework integrates state-of-the-art DNN architectures and requires minimal hyperparameters, avoiding the need for large validation sets.

Main Results:

• SSLDEC demonstrated superior performance compared to existing SSL methods on benchmark image classification tasks (MNIST, SVHN).
• Achieved 0.46% error on MNIST with 1000 labeled points and 4.43% error on SVHN with 500 labeled points.
• Significantly improved iso-intense infant brain MRI tissue segmentation compared to supervised-only training and pseudo-labeling methods.

Conclusions:

• SSLDEC effectively reduces the demand for costly expert annotations in machine learning.
• The framework is flexible and applicable to various DNN configurations for classification and segmentation.
• SSLDEC enhances the feasibility of automatic medical image segmentation and other data-scarce applications.