hlhong@swu.ac.kr

Our peritumoral and contextual patches can be applied to various image-based tumor analysis tasks as well as LC prediction of HCC in CT images. (This work was supported by Radiation Technology R&D program through the NRF of Korea (NRF-2017M2A2A7A02070427))

Prediction of tumor local control (LC) for hepatocellular carcinoma (HCC) patients provides important information in treatment planning. Several studies have reported that it is necessary to consider not only the information inside the tumor, but also the peritumoral information such as vascular invasion, and the context between the tumor and the organ, for predicting treatment responses. In this work, we propose an LC prediction method for HCC patients from CT images by integrating convolutional neural networks (CNNs) trained on multiple patches reflecting tumoral, peritumoral, and contextual information.

Our method was evaluated on a CT dataset acquired from 171 HCC patients with LC recordings. From the tumor ROI images, the tumoral, peritumoral, and contextual patches were generated to reflect the corresponding tumor characteristics. The four types of classifiers were then separately trained, including SVM trained on radiomic features, and three ResNets trained on each of the proposed patches. The four classifiers were then combined by score averaging to obtain an LC prediction. In experiments, the CNN trained on the peritumoral patches improved the sensitivity by 3.1%p compared to the conventional tumoral patch CNN, while the contextual patch CNN improved the specificity by 9.5%p compared to the tumoral patch CNN. The proposed method of combining multiple CNNs achieved the best performance with an accuracy of 77.19% and improved the accuracy by 9.94%p compared to the tumoral patch CNN.

The conventional tumoral patch CNN is efficient for learning the texture inside the tumor but has limitations in reflecting peritumoral information e.g. vascular invasion and contextual information e.g. tumor size. Throughout combining CNNs trained on the tumoral, peritumoral, and contextual patches, we confirmed that the proposed method improved the LC prediction accuracy by reflecting peritumoral and contextual information.