Increased interreader agreement in diagnosis of hepatocellular carcinoma using an adapted LI-RADS algorithm
详细信息 查看全文
- 作者：Anton S. Becker ; anton.becker@usz.ch ; Borna K. Barth ; Paulo H. Marquez ; Olivio F. Donati ; Erika J. Ulbrich ; Christoph Karlo ; Cä ; cilia S. Reiner ; Michael A. Fischer
- 关键词：MRI ; Hepatocellular carcinoma ; Liver ; LI-RADS
- 刊名：European Journal of Radiology
- 出版年：2017
- 出版时间：January 2017
- 年：2017
- 卷：86
- 期：Complete
- 页码：33-40
- 全文大小：1631 K
- 卷排序：86
文摘
To evaluate a simplified Liver Imaging Reporting and Data System (LI-RADS) algorithm to improve interreader agreement while maintaining diagnostic performance for HCC.Materials and methodsMRI scans of 84 cirrhotic patients with 104 distinct liver observations were retrospectively selected to equivocally match each of the LI-RADS grades (LR1-5) using histopathology and imaging follow up as standard of reference. Four independent radiologists categorized all observations as benign (LR1-2) or potentially malignant (LR3-5) and determined LI-RADS based imaging features including observation size, arterial phase hyperenhancement, washout, capsule appearance and threshold growth for LR3-5 observations and timed their readouts. LR3-5 observations were categorized according to the LI-RADS v2014 algorithm and according to a modified LI-RADS (mLI-RADS) version. Diagnostic performance and Interreader agreement were determined for LI-RADS and mLI-RADS using receiver operating characteristics (ROC) and Fleiss’ and Cohen’s Kappa analysis respectively.ResultsROC analysis revealed equal diagnostic performance for LI-RADS and mLI-RADS (area under the ROC curve = 0.91). Interreader agreement was higher using mLI-RADS as compared to current LI-RADS showing an improved overall (κ = 0.53 ± 0.04 vs. 0.45 ± 0.04), and pair-wise agreement between most readers (κ range 0.44-0.62 vs. 0.35-0.60) at a reduced median evaluation time (51 vs. 62 s per observation, p < 0.0001).ConclusionFocusing on observation size and washout criteria using a modified, stepwise LI-RADS decision tree for LR3-5 observations results in higher interobserver reliability and faster categorization while maintaining diagnostic accuracy.