Deep Semi-Supervised Algorithm for Learning Cluster-Oriented Representations of Medical Images Using Partially...

Area of Science:

• Medical Imaging Informatics
• Machine Learning
• Data Science

Background:

• Medical image datasets are challenging to organize due to non-standardized acquisition and storage.
• Manual data input errors by clinicians further complicate dataset homogeneity.
• Automated methods are needed for efficient and accurate medical image data extraction.

Purpose of the Study:

• To develop an algorithm for learning cluster-oriented representations of medical images.
• To fuse image data with partially observable DICOM tags for improved clustering.
• To enhance the accuracy of medical image dataset organization and retrieval.

Main Methods:

• Proposed a novel algorithm fusing medical images with partially observable DICOM tags.
• Modeled pairwise relations using the Gower distance measure across eight DICOM tags.
• Trained and tested models on large datasets (30,000 training, 8000 testing images) from a clinical PACS repository.
• Compared the proposed method against standard, deep unsupervised, and semi-supervised clustering algorithms.

Main Results:

• Achieved a Normalized Mutual Information (NMI) score of 0.584 for anatomic region clustering.
• Attained an NMI score of 0.793 for modality-based clustering.
• Demonstrated the effectiveness of using DICOM data for pairwise constraints in image clustering.

Conclusions:

• DICOM metadata, even when partially available, can significantly improve medical image clustering.
• The proposed algorithm offers a robust approach to organizing large medical image datasets.
• This method has the potential to enhance the usability of medical image archives for research and clinical applications.