Akazawa Kentaro, Sakamoto Ryo, Nakajima Satoshi, Wu Dan, Li Yue, Oishi Kenichi, Faria Andreia V, Yamada Kei, Togashi Kaori, Lyketsos Constantine G, Miller Michael I, Mori Susumu
Department of Radiology, Johns Hopkins University School of Medicine Baltimore, MD, United States.
Department of Radiology, Graduate School of Medical Science, Kyoto Prefectural University of Medicine Kyoto, Japan.
Front Neurol. 2019 Jan 24;10:7. doi: 10.3389/fneur.2019.00007. eCollection 2019.
To examine the feasibility and potential difficulties of automatically generating radiologic reports (RRs) to articulate the clinically important features of brain magnetic resonance (MR) images. We focused on examining brain atrophy by using magnetization-prepared rapid gradient-echo (MPRAGE) images. The technology was based on multi-atlas whole-brain segmentation that identified 283 structures, from which larger superstructures were created to represent the anatomic units most frequently used in RRs. Through two layers of data-reduction filters, based on anatomic and clinical knowledge, raw images (~10 MB) were converted to a few kilobytes of human-readable sentences. The tool was applied to images from 92 patients with memory problems, and the results were compared to RRs independently produced by three experienced radiologists. The mechanisms of disagreement were investigated to understand where machine-human interface succeeded or failed. The automatically generated sentences had low sensitivity (mean: 24.5%) and precision (mean: 24.9%) values; these were significantly lower than the inter-rater sensitivity (mean: 32.7%) and precision (mean: 32.2%) of the radiologists. The causes of disagreement were divided into six error categories: mismatch of anatomic definitions (7.2 ± 9.3%), data-reduction errors (11.4 ± 3.9%), translator errors (3.1 ± 3.1%), difference in the spatial extent of used anatomic terms (8.3 ± 6.7%), segmentation quality (9.8 ± 2.0%), and threshold for sentence-triggering (60.2 ± 16.3%). These error mechanisms raise interesting questions about the potential of automated report generation and the quality of image reading by humans. The most significant discrepancy between the human and automatically generated RRs was caused by the sentence-triggering threshold (the degree of abnormality), which was fixed to z-score >2.0 for the automated generation, while the thresholds by radiologists varied among different anatomical structures.
为检验自动生成放射学报告（RRs）以清晰阐述脑磁共振（MR）图像临床重要特征的可行性及潜在困难。我们聚焦于利用磁化准备快速梯度回波（MPRAGE）图像检查脑萎缩情况。该技术基于多图谱全脑分割，可识别283个结构，并据此创建更大的上层结构以代表RRs中最常用的解剖单元。通过基于解剖学和临床知识的两层数据缩减滤波器，原始图像（约10MB）被转换为几千字节的可读句子。该工具应用于92例有记忆问题患者的图像，并将结果与三位经验丰富的放射科医生独立生成的RRs进行比较。对分歧机制进行研究以了解人机界面在哪些方面成功或失败。自动生成的句子敏感性（平均：24.5%）和精确性（平均：24.9%）值较低；这些值显著低于放射科医生之间的敏感性（平均：32.7%）和精确性（平均：32.2%）。分歧原因分为六个错误类别：解剖定义不匹配（7.2±9.3%）、数据缩减错误（11.4±3.9%）、翻译错误（3.1±3.1%）、所用解剖术语空间范围差异（8.3±6.7%）、分割质量（9.8±2.0%）以及句子触发阈值（60.2±16.3%）。这些错误机制引发了关于自动报告生成潜力及人类图像解读质量的有趣问题。人工生成和自动生成的RRs之间最显著的差异是由句子触发阈值（异常程度）导致的，自动生成时该阈值固定为z分数>2.0，而放射科医生的阈值在不同解剖结构中有所不同。
