Development of deep learning model for prediction of chemotherapy response using PET images and radiomics features

Convolution neural network (CNN) was only consider their input images. It was only limitation of this technique for prediction of therapeutic response or patients outcome. The purpose of this study was to development of chemotherapy response prediction model using advanced deep learning techniques. 18F-FDG PET image was acquired in osteosarcoma patients (n=54) before the chemotherapy. The osteosarcoma patients was binary classification between responder (n=28) and non-responder (n=26). Radiomics feature was calculated using texture analysis method in 18F-FDG PET image. To improvement the deep learning model prediction accuracy, we performed the feature selection of information gain (IG) method. Deep learning models were define of CNN with radiomics features and without radiomics feature model. CNN was consist of 2 convolution layer and max pooling layers. Convolution and max pooling layers were generate the feature map. Advanced deep learning technique was input PET image as well as radiomics features. Radiomics features were concatenate with PET image feature map before the fully connected layer. Total 70 radiomics feature were extracted in 18F-FDG PET image of osteosarcoma patients before the chemotherapy. Total 9 radiomics features were selected by the above of the IG value 0.1. The prediction accuracy without radiomics feature model was 80%, On the other hand, the number of radiomics features , seventy, three, nine was 81, 82, 85% when they were worked with deep learning. mode. Nine radiomics features with deep learning model has better prediction accuracy than seventy radiomics feature. In this study, prediction accuracy of the selected radiomics features co-worked with deep learning model was higher than those of all features and only CNN model. Our developed prediction model will help to consider the radiomics feature as well as disease related feature such as genomics data.