Yongxin Zhang1,2, Xiaofei Lv3, Jiliang Qiu4, Bin Zhang1, Lu Zhang1, Jin Fang1, Minmin Li1, Luyan Chen1, Fei Wang1, Shuyi Liu1, Shuixing Zhang1. 1. Department of Radiology, The First Affiliated Hospital of Jinan University, Guangzhou, Guangdong, China. 2. Department of MR, Zhongshan City People's Hospital Affiliated to Sun Yat-sen University, Zhongshan, Guangdong, China. 3. Department of Medical Imaging, State Key Laboratory of Oncology in South China, Collaborative Innovation Center for Cancer Medicine, Sun Yat-Sen University Cancer Center, Guangzhou, China. 4. Department of Liver Surgery, State Key Laboratory of Oncology in South China, Collaborative Innovation Centre for Cancer Medicine, Sun Yat-sen University Cancer Centre, Guangzhou, China.
Abstract
BACKGROUND: Microvascular invasion (MVI) is a critical prognostic factor of hepatocellular carcinoma (HCC). However, it could only be obtained by postoperative histological examination. PURPOSE: To develop an end-to-end deep-learning models based on MRI images for preoperative prediction of MVI in HCC patients who underwent surgical resection. STUDY TYPE: Retrospective. POPULATION: Two hundred and thirty-seven patients with histologically confirmed HCC. FIELD STRENGTH: 1.5 T and 3.0 T. SEQUENCE: Axial T2 -weighted (T2 -w) with turbo spin echo sequence, T2 -Spectral Presaturation with Inversion Recovery (T2 -SPIR), and dynamic contrast-enhanced (DCE) imaging with fat suppressed enhanced T1 high-resolution isotropic volume examination. ASSESSMENT: The patients were randomly divided into training (N = 158) and validation (N = 79) sets. Data augmentation by random rotation was performed on the training set and the sample size increased to 1940 for each MR sequence. A three-dimensional convolutional neural network (3D CNN) was used to develop four deep-learning models, including three single-layer models based on single-sequence, and fusion model combining three sequences. MVI status was obtained from the postoperative pathology reports. STATISTICAL TESTS: The dice similarity coefficient (DSC) and Hausdorff distance (HD) were applied to assess the similarity and reproducibility between the manual segmentations of tumor from two radiologists. Receiver operating characteristic curve analysis was used to evaluate model performance. MVI was identified in 92 (38.8%) patients. Good reproducibility with interobserver DSCs of 0.90, 0.89, and 0.89 and HDs of 4.09, 3.67, and 3.60 was observed for PVP, T2 WI, and T2 -SPIR, respectively. The fusion model achieved an area under the curve (AUC) of 0.81, sensitivity of 69%, and specificity of 79% in the training set and 0.72, sensitivity of 55%, and specificity of 81% in the validation set. DATA CONCLUSION: 3D CNN model may serve as a noninvasive tool to predict MVI in HCC, whereas its accuracy needs to be enhanced with larger cohort. LEVEL OF EVIDENCE: 3 TECHNICAL EFFICACY: Stage 2.
BACKGROUND: Microvascular invasion (MVI) is a critical prognostic factor of hepatocellular carcinoma (HCC). However, it could only be obtained by postoperative histological examination. PURPOSE: To develop an end-to-end deep-learning models based on MRI images for preoperative prediction of MVI in HCC patients who underwent surgical resection. STUDY TYPE: Retrospective. POPULATION: Two hundred and thirty-seven patients with histologically confirmed HCC. FIELD STRENGTH: 1.5 T and 3.0 T. SEQUENCE: Axial T2 -weighted (T2 -w) with turbo spin echo sequence, T2 -Spectral Presaturation with Inversion Recovery (T2 -SPIR), and dynamic contrast-enhanced (DCE) imaging with fat suppressed enhanced T1 high-resolution isotropic volume examination. ASSESSMENT: The patients were randomly divided into training (N = 158) and validation (N = 79) sets. Data augmentation by random rotation was performed on the training set and the sample size increased to 1940 for each MR sequence. A three-dimensional convolutional neural network (3D CNN) was used to develop four deep-learning models, including three single-layer models based on single-sequence, and fusion model combining three sequences. MVI status was obtained from the postoperative pathology reports. STATISTICAL TESTS: The dice similarity coefficient (DSC) and Hausdorff distance (HD) were applied to assess the similarity and reproducibility between the manual segmentations of tumor from two radiologists. Receiver operating characteristic curve analysis was used to evaluate model performance. MVI was identified in 92 (38.8%) patients. Good reproducibility with interobserver DSCs of 0.90, 0.89, and 0.89 and HDs of 4.09, 3.67, and 3.60 was observed for PVP, T2 WI, and T2 -SPIR, respectively. The fusion model achieved an area under the curve (AUC) of 0.81, sensitivity of 69%, and specificity of 79% in the training set and 0.72, sensitivity of 55%, and specificity of 81% in the validation set. DATA CONCLUSION: 3D CNN model may serve as a noninvasive tool to predict MVI in HCC, whereas its accuracy needs to be enhanced with larger cohort. LEVEL OF EVIDENCE: 3 TECHNICAL EFFICACY: Stage 2.