Classification of Multi-Parametric Body MRI Series Using Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Multi-parametric magnetic resonance imaging (mpMRI) exams involve diverse series types and protocols.
• Inaccurate DICOM headers in mpMRI data hinder efficient radiologist review.
• A reliable method is needed to automatically classify mpMRI series types.

Purpose of the Study:

• To develop and evaluate a deep learning model for classifying 8 different body mpMRI series types.
• To compare the performance of various deep learning classifiers (ResNet, EfficientNet, DenseNet).
• To assess the model's performance with varying training data quantities and on external datasets.

Main Methods:

• Training and comparing ResNet, EfficientNet, and DenseNet classifiers on multi-institutional mpMRI data.
• Identifying the best-performing classifier (DenseNet-121) and evaluating its performance with different training data sizes.
• Testing the model on out-of-distribution datasets (DLDS, CPTAC-UCEC) and with different scanner data.

Main Results:

• DenseNet-121 achieved the highest F1-score (0.966) and accuracy (0.972) among the tested models.
• Accuracy exceeded 0.95 with over 729 training studies, improving with more data.
• The model demonstrated strong performance on external datasets, with accuracies of 0.872 (DLDS) and 0.810 (CPTAC-UCEC).

Conclusions:

• The DenseNet-121 model effectively classifies 8 body mpMRI series types with high accuracy.
• The model shows robustness and generalizability across internal and external datasets and different scanners.
• This deep learning approach enhances efficiency for radiologists interpreting mpMRI exams.