Highdicom: a Python Library for Standardized Encoding of Image Annotations and Machine Learning Model Outputs in...

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Pathology
• Radiology

Background:

• Machine learning (ML) shows promise in revolutionizing pathology and radiology diagnostics.
• Clinical integration of ML models is hindered by poor interoperability between ML systems and enterprise medical imaging systems.
• The DICOM standard, while comprehensive, is complex and poses challenges for the ML community.

Purpose of the Study:

• To present the highdicom library, a Python API simplifying the creation and parsing of DICOM-compliant files.
• To enable seamless integration of ML models with medical imaging systems.
• To facilitate the use of ML in pathology and radiology research and clinical practice.

Main Methods:

• Developed a high-level Python API (highdicom) abstracting low-level DICOM complexities.
• Leveraged NumPy arrays for efficient data representation within the Python ecosystem.
• Demonstrated highdicom's utility with slide microscopy and computed tomography imaging datasets.

Main Results:

• highdicom enables encoding and decoding of image-derived information in DICOM format with minimal Python code.
• The library ensures interoperability between ML workflows and clinical imaging systems.
• Experiments confirmed highdicom's effectiveness in training and evaluating ML models compliant with DICOM standards.

Conclusions:

• highdicom bridges the gap between ML research and clinical medical imaging systems.
• The library promotes standardization and streamlines ML model development and deployment in pathology and radiology.
• highdicom is available as a free and open-source tool to foster wider adoption.