Abdominal and Pelvic MRI Protocol Prediction Using Natural Language Processing

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Radiology Informatics

Background:

• Radiologists face significant workload from non-interpretive tasks like exam protocoling.
• Automating protocoling can enhance departmental efficiency and reduce radiologist burden.

Purpose of the Study:

• To develop and evaluate a natural language processing (NLP) artificial intelligence (AI) solution for automated protocoling of abdomen and pelvis MRI exams.
• To assess the AI model's performance using historical patient metadata and order information.

Main Methods:

• A retrospective study utilizing de-identified metadata from approximately 46,000 adult abdomen and pelvis MRI exams (2019-2021).
• Fine-tuning a Bidirectional Encoder Representations from Transformers (BERT) NLP model in sequence classification mode.
• Excluding 12 months of data during the COVID pandemic to mitigate bias.

Main Results:

• The trained AI model achieved an accuracy of 88.5% with a Matthews correlation coefficient of 0.874.
• Expert review of model errors revealed 81.9% were correct or reasonable alternative protocols, yielding a real-world accuracy of 97.9%.

Conclusions:

• NLP algorithms, including BERT-based models, can effectively predict MRI imaging protocols for the abdomen and pelvis.
• This AI solution has the potential to significantly decrease non-interpretive task load for radiologists and improve overall departmental efficiency.