AUTOMATED CLASSIFICATION OF MULTI-PARAMETRIC BODY MRI SERIES

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Multi-parametric MRI (mpMRI) is crucial for disease diagnosis, but DICOM header inaccuracies hinder automated series arrangement.
• Current practices require manual correction of mpMRI series for radiologist hanging protocols, increasing workload.

Purpose of the Study:

• To develop and validate an automated framework for classifying 8 different series types within mpMRI studies.
• To enhance the efficiency and accuracy of radiology hanging protocols through automated series classification.

Main Methods:

• A DenseNet-121 model was trained using 1,363 mpMRI studies from three Siemens scanners.
• 5-fold cross-validation was employed for model training.
• The model's performance was evaluated on a separate test set of 313 mpMRI studies.

Main Results:

• The automated framework achieved high performance metrics: 96.6% average precision, 96.6% sensitivity, 99.6% specificity, and 96.6% F1 score.
• The method demonstrated robust classification of MRI series for chest, abdomen, and pelvis imaging.

Conclusions:

• This pilot work presents the first method for classifying mpMRI series types in chest, abdomen, and pelvis imaging.
• The proposed automated framework offers a robust solution for automating hanging protocols in clinical radiology.
• The findings suggest significant potential for improving workflow efficiency and reducing errors in modern radiology practices.