Uncertainty-aware automatic TNM staging classification for [¹⁸F] Fluorodeoxyglucose PET-CT reports for lung cancer...

Area of Science:

• Medical Imaging and Oncology
• Artificial Intelligence in Healthcare
• Natural Language Processing for Clinical Data

Background:

• [18F] Fluorodeoxyglucose (FDG) PET-CT is crucial for lung cancer diagnosis and staging.
• Clinical findings are in free-text reports, requiring manual expert categorization.
• Pre-trained language models (PLMs) show promise in extracting complex linguistic features from medical texts.

Purpose of the Study:

• To develop a multi-task classifier ('TNMu') for extracting Tumor, Node, Metastasis (TNM) staging information from lung cancer PET-CT reports.
• To incorporate an uncertainty classification task ('u') for ambiguous TNM statuses.
• To evaluate the effectiveness of different PLMs and machine learning approaches for this task.

Main Methods:

• Annotated 2498 PET-CT reports for TNM staging and uncertainty.
• Trained and evaluated eleven publicly available PLMs.
• Compared multi-task, single-task, and traditional machine learning models.
• Validated performance on an internal dataset and an external dataset of 461 reports.

Main Results:

• The multi-task 'TNMu' classifier using GatorTron achieved 84% overall accuracy on the internal test set and 79% on the external set.
• Individual TNM task performance approached expert levels with macro average F1 scores of 0.91, 0.95, and 0.90 on external data.
• An F1 score of 0.77 was achieved for the uncertainty classification task.

Conclusions:

• The 'TNMu' classifier effectively extracts TNM staging information from PET-CT reports, demonstrating strong performance on both internal and external datasets.
• Multi-task approaches offer superior performance and computational efficiency compared to single-task PLMs.
• This AI tool can enhance PET-CT services by assisting in auditing, research cohort creation, and developing clinical decision support systems.