Image features related to margin and enhancement pattern improve the performance of computer-aided diagnosis for hepatic diseases using multi-phase computed tomography

Background This study aimed to develop features related to the lesion margin and enhancement pattern, which are very important in the radiologic diagnostic process. We also aimed to implement and investigate these features in the computer-aided diagnosis (CAD) of hepatic diseases using computed tomography (CT). Methods We retrospectively analyzed 378 lesions with 1 512 multi-phase CT images of liver lesions. We used ensemble methods to create classification models. Two types of features were developed and used as predictors, namely, margin features and relative spatial intensity ratio (RSIR) features. Margin features were extracted using Gabor transformation and the sigmoid function whereas RSIR features were obtained by calculating the concentration and distribution of the contrast in the lesion against the surrounding hepatic parenchyma. To assess these two types of features and compare them with other features used in previous studies, we created models for multi-class classification using different feature subsets. Accuracy, kappa, and AUC were calculated. The importance and interactions of predictors were also estimated. Results The classification model with margin features exhibited the best performance (accuracy: 0.89 +/- 0.04; kappa: 0.85 +/- 0.06), followed by that with RISR features (accuracy: 0.85 +/- 0.05; kappa: 0.79 +/- 0.07). The plots for variable importance and interactions also showed these two types of features were important in classification models and that they interacted with other features. Conclusions Lesion margin and enhancement pattern are helpful in CAD. The features we have developed are general and can be easily adapted to other diagnostic scenarios in which CT and other imaging modalities are used.