Automated identification of MRI series using a hierarchical modular machine-learning pipeline.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology
• Machine Learning for Healthcare

Background:

• Manual annotation of large Magnetic Resonance Imaging (MRI) datasets is time-consuming and costly.
• Existing methods relying solely on DICOM headers are unreliable due to data heterogeneity and potential inaccuracies.
• Automated labeling is crucial for efficient cataloging and analysis of extensive MR imaging data.

Purpose of the Study:

• To develop and evaluate an AI-based modular model for automated classification of MR imaging series.
• To address the limitations of manual annotation and unreliable DICOM header information.
• To provide a scalable solution for labeling large and diverse MR imaging datasets.

Main Methods:

• A five-sequential-classifier pipeline (Family, Weighting, Fat Suppression, Contrast, Others) was developed.
• The model was trained and tested on 18,181 MR imaging series from the multicenter PRIMAGE repository.
• A hybrid approach combining DICOM tag-based machine learning (CatBoost/Random Forest) and image analysis (ResNet-50 for contrast) was employed.

Main Results:

• High classification accuracy was achieved for Weighting (0.994), Family (0.984), Fat Suppression (0.959), and Others (0.958).
• The Contrast classifier achieved an accuracy of 0.841.
• The overall end-to-end classification yielded a weighted F1 score of 0.849 and an accuracy of 0.853.

Conclusions:

• The AI pipeline offers a reliable and scalable method for labeling heterogeneous MR imaging datasets.
• The model demonstrates excellent performance in classifying MR series, significantly reducing manual curation efforts.
• Contrast classification remains an area for improvement, requiring further refinement or additional modules.