| Popis: |
Radiomics is a powerful tool in imaging-based clinical studies and it is gaining importance in oncology due to its ability to extract high-throughput image-based biomarkers in a non-invasive way. However, common methods for extracting radiomics features are mainly designed in the CPU, being an extremely time-consuming process when complicated method verifications or large datasets are involved. In this study, a deep learning-based radiomics sequencer was developed that is capable to extract a total of 93 features, including all histogram and texture based features. Our proposed network exploits the advantage of current CPU-based radiomics feature extraction methods because it has the potential to maintain the interpretability of radiomics features while reducing considerably the computation time of the feature extraction process. To illustrate the effectiveness of the network, we perform the experiments on BRATS17 image dataset obtaining the mean squared error (MSE) of 0.00092 in training, 0.0023 in validation and 0.00197 in test datasets. The feature extraction process was completed in 22.21 seconds, being 7205% faster than conventional CPU-based method. In this way, the potential advantages provided by the proposed network allow radiomics feature extraction process to be faster and more adaptive. |
