Deep multi-task learning and random forest for series classification by pulse sequence type and orientation

Area of Science:

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

Background:

• Rule-based MRI image routing is limited by complex studies and naming variations.
• Accurate series identification is crucial for efficient post-processing and Picture Archiving and Communication System (PACS) viewing.
• Existing deep learning models classify only basic brain MRI sequences.

Purpose of the Study:

• To develop and validate an ensemble model for classifying 25 brain MRI sequences and image orientation.
• To address limitations of rule-based routing in complex, multi-site healthcare environments.
• To improve automated image identification for enhanced post-processing and PACS workflows.

Main Methods:

• An ensemble model combining a 2D convolutional neural network and a random forest classifier was developed.
• The model was trained on DICOM metadata to classify series by sequence and orientation.
• Two datasets were used: Dataset A (institutional) and Dataset B (external validation) with extensive series and images.

Main Results:

• The ensemble model achieved 98% overall sequence accuracy on the institutional dataset (Dataset A).
• The ensemble model achieved 99% overall sequence accuracy on the external validation dataset (Dataset B).
• All models demonstrated over 99% accuracy in classifying image orientation on both datasets.

Conclusions:

• The developed ensemble model effectively handles the complexity of brain MRI studies in clinical practice.
• This approach offers a more comprehensive solution than previous methods, classifying a wider range of sequences and orientations.
• The model enhances automated series identification, supporting advanced clinical workflows.