Sequence-Type Classification of Brain MRI for Acute Stroke Using a Self-Supervised Machine Learning Algorithm.

Area of Science:

• Medical Imaging
• Machine Learning
• Radiology

Background:

• Accurate classification of brain MRI sequences is crucial for diagnosis and treatment.
• Current methods may rely on manual labeling, which can be time-consuming and prone to error.
• Developing automated, reliable classification systems is a key challenge in medical imaging analysis.

Purpose of the Study:

• To propose a self-supervised machine learning algorithm for brain MRI sequence-type classification.
• To utilize DICOM metadata as a supervisory signal for training.
• To develop a sustainable self-learning system for automated MRI classification.

Main Methods:

• Developed ImageSort-net, a machine learning framework utilizing MRI acquisition parameters.
• Created rule-based virtual labels from DICOM metadata for training.
• Trained and evaluated models using hospital and multi-center trial datasets, comparing ML algorithms trained with virtual labels (MLvirtual) and human expert labels (MLhumans).

Main Results:

• ImageSort-net (MLvirtual) demonstrated comparable accuracy to MLhumans (98.5% vs. 99%) on hospital datasets.
• On smaller multi-center datasets, MLvirtual showed lower accuracy (95.6% vs. 99.4%) but improved significantly after retraining with integrated data (99.7%).
• Retrained MLvirtual and MLhumans achieved identical inference performance on multi-center datasets (99.7%).

Conclusions:

• Self-supervised machine learning using rule-based virtual labels from DICOM metadata is effective for brain MRI sequence classification.
• The ImageSort-net framework offers a sustainable self-learning system for medical imaging.
• This approach reduces reliance on manual labeling and enhances classification accuracy, particularly when integrating diverse datasets.