Automated Classification of Body MRI Sequences Using Convolutional Neural Networks.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Standardized naming conventions for MRI protocols are lacking, causing inconsistencies in DICOM headers.
• Variations in MRI scanners, practices, and technologist preferences lead to conflicts in series descriptions.
• These inconsistencies impact hanging protocols and require manual clinician oversight for accurate diagnosis.

Purpose of the Study:

• To develop and evaluate a classification model for five different series in multi-parametric MRI (mpMRI) studies.
• To address the challenges posed by non-standardized MRI protocols in clinical practice.
• To enable automated hanging protocols and facilitate large-scale data cohort creation for research.

Main Methods:

• Comparison of 2D and 3D classification networks (ResNet-50, ResNet-101, DenseNet-121, EfficientNet-BN0) using Siemens scanner data.
• Analysis of model performance with varying training data quantities and data augmentation.
• Evaluation of out-of-distribution (OOD) robustness on Philips scanner data and testing with downsampled/cropped inputs.
• Training the optimal model on combined Siemens and Philips scanner data to improve cross-scanner performance.

Main Results:

• The 3D DenseNet-121 ensemble achieved a 99.5% F1 score on Siemens data and 86.5% on Philips OOD data.
• Data augmentation and input size reduction did not significantly impact model performance.
• Training on combined data improved performance to 98.8% for Philips and 99.3% for Siemens test sets.
• The 3D DenseNet-121 model demonstrated robust performance across different scanner types.

Conclusions:

• The developed model effectively classifies mpMRI sequences in chest, abdomen, and pelvis studies.
• This approach offers potential for robust automation of hanging protocols.
• The classification method can aid in creating large-scale data cohorts for pre-clinical research.