Deep Learning-driven classification of external DICOM studies for PACS archiving.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology Informatics

Background:

• Patients frequently transfer between hospitals, necessitating the import of external imaging studies into new Picture Archiving and Communication Systems (PACS).
• Manual mapping of these studies is time-consuming and prone to errors, impacting patient care and data management.

Purpose of the Study:

• To present MOdality Mapping and Orchestration (MOMO), a novel deep learning-based approach for automating the classification and mapping of external DICOM imaging studies.
• To evaluate MOMO's performance against existing methods and a commercial product for accuracy and efficiency in PACS archiving.

Main Methods:

• A deep learning ensemble, including DenseNet-161 and ResNet-152, was trained on 11,934 local imaging series with anatomical labels.
• A custom algorithm was developed to extract and integrate imaging metadata with the neural network ensemble for robust study classification.
• Performance was assessed using 843 anonymized external studies, with several algorithm variations tested.

Main Results:

• MOMO achieved 92.71% accuracy with 2.63% minor errors, significantly outperforming a commercial product (82.86% accuracy) and a pure neural network ensemble (72.69% accuracy).
• The optimized algorithm, combining all information into a single vote-based classifier, demonstrated the highest performance.
• The system successfully identified 76 medical study types across seven modalities, including CT, MRI, and ultrasound.

Conclusions:

• Deep learning combined with metadata analysis offers a flexible and effective solution for automated DICOM study classification in PACS.
• MOMO provides a significant accuracy improvement over current commercial solutions, enhancing the efficiency of medical image data management.
• The continuous application of deep learning techniques promises further advancements in automated medical image classification and archiving.