Natural Language Processing for Imaging Protocol Assignment: Machine Learning for Multiclass Classification of...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Natural Language Processing

Background:

• Accurate protocol assignment is crucial for high-quality CT imaging.
• Automating this process using NLP can enhance efficiency and accuracy.
• Abdominal CT protocol assignment is complex due to numerous indications and parameters.

Purpose of the Study:

• To compare the performance of conventional machine learning, deep learning, and automated machine learning (AutoML) for multiclass text classification of abdominal CT protocols.
• To evaluate different algorithms including Random Forest, Tree Ensemble, Gradient Boosted Tree, multi-layer perceptron, and ULMFiT.

Main Methods:

• A dataset of 94,501 abdominal CT studies and their assigned protocols was analyzed.
• Free text indications and ICD codes were used for NLP analysis and protocol prediction.
• Algorithms were tested on both unbalanced and balanced datasets, with data augmentation applied.

Main Results:

• On unbalanced data, AutoML achieved the highest F1 score (0.854), outperforming manual ML and deep learning models.
• On balanced data, the Tree Ensemble algorithm performed best (F1 score 0.803).
• AutoML required longer training/evaluation times (4h 45min) compared to manual methods (average 51min).

Conclusions:

• Machine learning and NLP are effective for the complex task of abdominal CT protocol assignment.
• AutoML demonstrates strong performance for this application, despite longer processing times.
• Further research into optimizing automated protocol assignment can improve imaging quality and efficiency.