Semi-supervised Machine Learning with MixMatch and Equivalence Classes.

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Machine Learning

Background:

• Semi-supervised learning (SSL) methods are crucial for leveraging large datasets with scarce labels in computer vision.
• Existing SSL approaches have not been extensively applied to medical imaging tasks.
• MixMatch and Nullspace Tuning are promising SSL techniques for improving performance with limited labeled data.

Purpose of the Study:

• To explore the application of MixMatch combined with Nullspace Tuning in medical imaging.
• To characterize the impact of these methods on model performance with diminishing labeled data.
• To evaluate the efficacy of these techniques in skin lesion diagnosis and lung cancer prediction.

Main Methods:

• Implementation of MixMatch, a semi-supervised learning algorithm, adapted for medical imaging.
• Integration of Nullspace Tuning, a method for leveraging unlabeled data from equivalence classes.
• Evaluation of supervised, MixMatch, Nullspace Tuning, and combined MixMatch-Nullspace Tuning models on skin lesion and lung cancer datasets with varying amounts of labeled data.

Main Results:

• The combined MixMatch with Nullspace Tuning achieved an AUC of 0.755 for lung cancer prediction using only 200 labeled subjects.
• For skin lesion diagnosis, the combined method reached a balanced multi-class accuracy of 77% with 779 labeled examples.
• Performance with limited labels using the combined approach was comparable to fully supervised methods utilizing all available labels.

Conclusions:

• MixMatch with Nullspace Tuning demonstrates significant potential for advancing semi-supervised learning in medical imaging.
• These methods effectively mitigate challenges associated with limited data acquisition and annotation in medical AI.
• Integrating state-of-the-art SSL techniques is vital for data-driven medical imaging advancements.