Classification of computed tomography scans: a novel approach implementing an enforced random forest algorithm.

Area of Science:

• Medical Imaging
• Radiology
• Machine Learning in Healthcare

Background:

• Medical imaging, particularly Computed Tomography (CT), faces challenges in radiation dose management and protocol standardization.
• Current methods for protocol classification are manual and time-consuming.
• There is a need for data-driven solutions to optimize CT protocols and ensure patient safety.

Purpose of the Study:

• To introduce a machine learning approach for classifying CT scan protocols.
• To leverage dose monitoring data for automated protocol categorization.
• To provide a data-driven solution for establishing Diagnostic Reference Levels (DRLs) efficiently.

Main Methods:

• Developed a classification workflow using a Random Forest Classifier.
• Categorized CT scans into anatomical regions (head, thorax, abdomen, spine, complex multi-region).
• Employed an iterative "human-in-the-loop" refinement process including expert validation.

Main Results:

• Analyzed 52,982 CT scan records from 11 imaging devices across five hospitals.
• Trained a classifier to distinguish multiple anatomical regions with high accuracy.
• Achieved 97% accuracy in final validation on an independent dataset, demonstrating model robustness.

Conclusions:

• Introduced a novel, data-driven approach for medical imaging protocol classification using a random forest algorithm.
• Created a framework for protocol classification and DRL establishment by integrating computational intelligence with clinical expertise.
• Demonstrated the potential of machine learning to transform CT protocol management and pave the way for application in other radiological procedures.