Wu Zhou1, Guangyi Wang2, Guoxi Xie3, Lijuan Zhang4. 1. School of Medical Information Engineering, Guangzhou University of Chinese Medicine, Guangzhou, China, 510006. 2. Department of Radiology, Guangdong General Hospital, Guangzhou, China, 510080. 3. School of Basic Medical Sciences, Guangzhou Medical University, Guangzhou, China, 510182. 4. Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, 510085.
Abstract
PURPOSE: To effectively grade hepatocellular carcinoma (HCC) based on deep features derived from diffusion weighted images (DWI) with multiple b-values using convolutional neural networks (CNN). MATERIALS AND METHODS: Ninety-eight subjects with 100 pathologically confirmed HCC lesions from July 2012 to October 2018 were included in this retrospective study, including 47 low-grade and 53 high-grade HCCs. DWI was performed for each subject with a 3.0T MR scanner in a breath-hold routine with three b-values (0,100, and 600 s/mm2 ). First, logarithmic transformation was performed on original DWI images to generate log maps (logb0, logb100, and logb600). Then, a resampling method was performed to extract multiple 2D axial planes of HCCs from the log map to increase the dataset for training. Subsequently, 2D CNN was used to extract deep features of the log map for HCCs. Finally, fusion of deep features derived from three b-value log maps was conducted for HCC malignancy classification. Specifically, a deeply supervised loss function was devised to further improve the performance of lesion characterization. The data set was split into two parts: the training and validation set (60 HCCs) and the fixed test set (40 HCCs). Four-fold cross validation with 10 repetitions was performed to assess the performance of deep features extracted from single b-value images for HCC grading using the training and validation set. Receiver operating characteristic curve (ROC) and area under the curve (AUC) values were used to assess the characterization performance of the proposed deep feature fusion method to differentiate low-grade and high-grade in the fixed test set. RESULTS: The proposed fusion of deep features derived from logb0, logb100, and logb600 with deeply supervised loss function generated the highest accuracy for HCC grading (80%), thus outperforming the method of deep feature derived from the ADC map directly (72.5%), the original b0 (65%), b100 (68%), and b600 (70%) images. Furthermore, AUC values of the deep features of the ADC map, the deep feature fusion with concatenation, and the proposed deep feature fusion with deeply supervised loss function were 0.73, 0.78, and 0.83, respectively. CONCLUSION: The proposed fusion of deep features derived from the logarithm of the three b-value images yields high performance for HCC grading, thus providing a promising approach for the assessment of DWI in lesion characterization.
PURPOSE: To effectively grade hepatocellular carcinoma (HCC) based on deep features derived from diffusion weighted images (DWI) with multiple b-values using convolutional neural networks (CNN). MATERIALS AND METHODS: Ninety-eight subjects with 100 pathologically confirmed HCC lesions from July 2012 to October 2018 were included in this retrospective study, including 47 low-grade and 53 high-grade HCCs. DWI was performed for each subject with a 3.0T MR scanner in a breath-hold routine with three b-values (0,100, and 600 s/mm2 ). First, logarithmic transformation was performed on original DWI images to generate log maps (logb0, logb100, and logb600). Then, a resampling method was performed to extract multiple 2D axial planes of HCCs from the log map to increase the dataset for training. Subsequently, 2D CNN was used to extract deep features of the log map for HCCs. Finally, fusion of deep features derived from three b-value log maps was conducted for HCC malignancy classification. Specifically, a deeply supervised loss function was devised to further improve the performance of lesion characterization. The data set was split into two parts: the training and validation set (60 HCCs) and the fixed test set (40 HCCs). Four-fold cross validation with 10 repetitions was performed to assess the performance of deep features extracted from single b-value images for HCC grading using the training and validation set. Receiver operating characteristic curve (ROC) and area under the curve (AUC) values were used to assess the characterization performance of the proposed deep feature fusion method to differentiate low-grade and high-grade in the fixed test set. RESULTS: The proposed fusion of deep features derived from logb0, logb100, and logb600 with deeply supervised loss function generated the highest accuracy for HCC grading (80%), thus outperforming the method of deep feature derived from the ADC map directly (72.5%), the original b0 (65%), b100 (68%), and b600 (70%) images. Furthermore, AUC values of the deep features of the ADC map, the deep feature fusion with concatenation, and the proposed deep feature fusion with deeply supervised loss function were 0.73, 0.78, and 0.83, respectively. CONCLUSION: The proposed fusion of deep features derived from the logarithm of the three b-value images yields high performance for HCC grading, thus providing a promising approach for the assessment of DWI in lesion characterization.