Dataset Information

Deep convolutional neural network applied to the liver imaging reporting and data system (LI-RADS) version 2014 category classification: a pilot study.

ABSTRACT:

Purpose

To develop a deep convolutional neural network (CNN) model to categorize multiphase CT and MRI liver observations using the liver imaging reporting and data system (LI-RADS) (version 2014).

Methods

A pre-existing dataset comprising 314 hepatic observations (163 CT, 151 MRI) with corresponding diameters and LI-RADS categories (LR-1-5) assigned in consensus by two LI-RADS steering committee members was used to develop two CNNs: pre-trained network with an input of triple-phase images (training with transfer learning) and custom-made network with an input of quadruple-phase images (training from scratch). The dataset was randomly split into training, validation, and internal test sets (70:15:15 split). The overall accuracy and area under receiver operating characteristic curve (AUROC) were assessed for categorizing LR-1/2, LR-3, LR-4, and LR-5. External validation was performed for the model with the better performance on the internal test set using two external datasets (EXT-CT and EXT-MR: 68 and 44 observations, respectively).

Results

The transfer learning model outperformed the custom-made model: overall accuracy of 60.4% and AUROCs of 0.85, 0.90, 0.63, 0.82 for LR-1/2, LR-3, LR-4, LR-5, respectively. On EXT-CT, the model had an overall accuracy of 41.2% and AUROCs of 0.70, 0.66, 0.60, 0.76 for LR-1/2, LR-3, LR-4, LR-5, respectively. On EXT-MR, the model had an overall accuracy of 47.7% and AUROCs of 0.88, 0.74, 0.69, 0.79 for LR-1/2, LR-3, LR-4, LR-5, respectively.

Conclusion

Our study shows the feasibility of CNN for assigning LI-RADS categories from a relatively small dataset but highlights the challenges of model development and validation.

SUBMITTER: Yamashita R

PROVIDER: S-EPMC6946904 | biostudies-literature | 2020 Jan

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Deep convolutional neural network applied to the liver imaging reporting and data system (LI-RADS) version 2014 category classification: a pilot study.

Yamashita Rikiya R Mittendorf Amber A Zhu Zhe Z Fowler Kathryn J KJ Santillan Cynthia S CS Sirlin Claude B CB Bashir Mustafa R MR Do Richard K G RKG

Abdominal radiology (New York) 20200101 1

<h4>Purpose</h4>To develop a deep convolutional neural network (CNN) model to categorize multiphase CT and MRI liver observations using the liver imaging reporting and data system (LI-RADS) (version 2014).<h4>Methods</h4>A pre-existing dataset comprising 314 hepatic observations (163 CT, 151 MRI) with corresponding diameters and LI-RADS categories (LR-1-5) assigned in consensus by two LI-RADS steering committee members was used to develop two CNNs: pre-trained network with an input of triple-pha ...[more]

PMID: 31696269

Similar Datasets

Project description:ObjectivesThis study assesses the risk of progression of Liver Imaging Reporting and Data System (LI-RADS) categories, and the effects of inter-exam changes in modality or radiologist on LI-RADS categorization.MethodsClinical LI-RADS v2014 CT and MRI exams at our institution between January 2014 and September 2017 were retrospectively identified. Untreated LR-1, LR-2, LR-3, and LR-4 observations with at least one follow-up exam were included. Three hundred and seventy-two observations in 214 patients (149 male, 65 female, mean age 61 ± 10 years) were included during the study period (715 exams total). Cumulative incidence curves for progression to malignant LI-RADS categories (LR-5 or LR-M) and to LR-4 or higher were generated for each index category and compared using log-rank tests with a resampling extension. Relationships between inter-exam changes in LI-RADS category and modality or radiologist, adjusted for inter-exam time intervals, were modeled using mixed effect logistic regressions.ResultsMedian inter-exam follow-up interval and total follow-up duration were 123 and 227 days, respectively. Index LR-1, LR-2, LR-3, and LR-4 differed significantly in their cumulative incidences of progression to malignant categories (p < 0.0001), which were 0%, 2%, 7%, and 32% at 6 months, respectively. Index LR-1, LR-2, and LR-3 differed significantly in cumulative incidences of progression to LR-4 or higher (p = 0.003). MRI-MRI exam pairs had more stable LI-RADS categorization compared to CT-CT (OR = 0.460, p = 0.0018).ConclusionsLI-RADS observations demonstrate increasing risk of progression to malignancy with increasing category ranging from 0% for LR-1 to 32% for LR-4 at 6 months. Inter-exam modality changes are associated with LI-RADS category changes.Key points• While the majority of LR-2 observations remain stable over long-term follow-up, LR-3 and especially LR-4 observations have a higher risk for category progression. • Category transitions between sequential exams using different modalities (CT vs. MRI) may reflect modality differences rather than biological change. MRI, especially with the same type of contrast agent, may provide the most reproducible categorization, although this needs additional validation. • In a clinical practice setting, in which radiologists refer to prior imaging and reports, there was no significant association between changes in radiologist and changes in LI-RADS categorization.

Project description:ObjectivesTo explore the role of preoperative gadoxetic acid-enhanced MRI in stratifying the risk of early recurrence in patients with LR-5 hepatocellular carcinoma (HCC) by LI-RADS v2018 after curative resection.MethodsBetween July 2015 and August 2018, this study evaluated consecutive treatment-naïve at-risk LR-5 HCC patients who underwent gadoxetic acid-enhanced MRI examination within 2 weeks before curative resection. The Cox regression analysis was performed to identify potential predictors of early recurrence. Disease-free survival (DFS) rates were analyzed and compared by using the Kaplan-Meier method and log-rank tests.ResultsFifty-three of 103 (51.5%) patients experienced early recurrence. Three MRI findings were significantly associated with early recurrence: corona enhancement (hazard ratio [HR]: 2.116; p = 0.013), peritumoral hypointensity on hepatobiliary phase (HBP) (HR: 2.262; p = 0.007), and satellite nodule (HR: 2.777; p = 0.005). An additional risk factor was AFP level > 400 ng/mL (HR: 1.975; p = 0.016). Based on the number of MRI predictors, LR-5 HCC patients were stratified into three subgroups: LR-5a (60/103; no predictor), LR-5b (26/103; one predictor), and LR-5c (17/103; two or three predictors), with low, medium, and high risk of early recurrence, respectively. The 2-year DFS rate of LR-5a, LR-5b, and LR-5c patients was 65.0%, 38.5%, and 5.9%, respectively, while the corresponding median DFS was undefined, 17.1 months, and 5.1 months, respectively (p < 0.001).ConclusionsIn at-risk LR-5 HCC patients, corona enhancement, peritumoral hypointensity on HBP, and satellite nodule could be used to preoperatively stratify the risk of early recurrence after hepatectomy.Key points• Corona enhancement, peritumoral hypointensity on HBP, satellite nodule, and serum AFP level > 400 ng/mL were significant predictors of early recurrence in patients with LR-5 HCC after hepatectomy. • Based on the number of predictive MRI findings, LR-5 HCC patients could be preoperatively stratified into three subgroups: LR-5a, LR-5b, and LR-5c, with significantly different risk of early recurrence and disease-free survival. • Preoperative risk stratification is essential for the identification of patients at increased risk of postoperative early recurrence, which may contribute to risk-based personalized management for LR-5 HCC patients.

Dataset Information

Deep convolutional neural network applied to the liver imaging reporting and data system (LI-RADS) version 2014 category classification: a pilot study.

Purpose

Methods

Results

Conclusion

Publications

Deep convolutional neural network applied to the liver imaging reporting and data system (LI-RADS) version 2014 category classification: a pilot study.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets