基于深度学习的乳腺超声报告 BI-RADS 抽取增强型自然语言处理方法

摘要

乳腺癌是全球致死率最高的疾病之一，准确解读乳腺超声报告对于早期诊断与治疗至关重要。然而，这类报告中的关键发现往往埋没在非结构化文本里，给自动化抽取造成困难。本文提出一种基于深度学习的自然语言处理（NLP）方法，用于从乳腺超声数据中抽取乳腺影像报告与数据系统（BI-RADS）类别。我们在一批取自沙特阿拉伯某医院、经人工标注的报告数据集上训练了循环神经网络（RNN）模型，具体采用双向长短期记忆网络（BiLSTM）结构；同时引入不确定性估计技术来处理歧义病例，并借助数据增强提升模型性能。实验结果表明，本文的深度学习方法优于传统的基于规则的方法与机器学习方法，在分类任务中取得了令人满意的准确率。本研究对于放射学报告的自动化、辅助临床决策以及推进乳腺癌研究均具有重要意义。

Abstract

Breast cancer stands as one of the top causes of death around the globe, making the accurate interpretation of breast ultrasound reports vital for early diagnosis and treatment. Unfortunately, key findings in these reports are often buried in unstructured text, complicating automated extraction. This study presents a deep learning-based natural language processing (NLP) approach to extract breast imaging reporting and data system (BI-RADS) categories from breast ultrasound data. We trained a recurrent neural network (RNN) model, specifically using a BiLSTM architecture, on a dataset of reports that were manually annotated from a hospital in Saudi Arabia. Our approach also incorporates uncertainty estimation techniques to tackle ambiguous cases and uses data augmentation to boost model performance. The experimental results indicate that our deep learning method surpasses traditional rule-based and machine-learning techniques, achieving impressive accuracy in classification tasks. This research plays a significant role in automating radiology reporting, aiding clinical decision-making, and pushing forward the field of breast cancer research.