Natural Language Generation Model for Mammography Reports Simulation.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Natural Language Processing

Background:

• Training machine learning algorithms requires large labeled medical corpora.
• Data augmentation through text report simulation is crucial for expanding datasets.
• Medical text generation faces challenges in maintaining report realism, patient privacy, and clinical accuracy.

Purpose of the Study:

• To develop a conditioned LSTM-RNN architecture for simulating realistic mammography reports.
• To evaluate the quality and utility of simulated reports for AI model training and analysis.
• To compare the performance of the proposed generative model against traditional methods like Markov Random Fields.

Main Methods:

• Utilized a conditioned Long Short-Term Memory-Recurrent Neural Network (LSTM-RNN) architecture.
• Generated synthetic mammography reports, preserving clinical content and stylistic elements.
• Evaluated simulated reports through content analysis, classification tasks (benign/malignant), and radiologist assessment.
• Compared performance metrics against Markov Random Fields (MRFs).

Main Results:

• Simulated reports demonstrated realistic content and style, with a radiologist classifying 75% as authentic.
• The LSTM-RNN model achieved strong performance in classifying simulated reports into benign and malignant classes, with a high average AUC.
• The RNN-LSTM generative model significantly outperformed MRFs in terms of performance and stability.

Conclusions:

• The conditioned LSTM-RNN model effectively generates realistic mammography reports suitable for data augmentation.
• This approach enhances the potential for training more robust AI models in medical imaging analysis.
• The simulated reports show promise for improving diagnostic accuracy and medical research without compromising patient privacy.