Purpose: Our aim was to predict the recurrence after radiation therapy by using local binary pattern (LBP)-based dosiomics for head and neck squamous cell cancer (HNSCC) patients.
Methods: A total of 131 HNSCC patients has recurrence/non-recurrence information after radiation therapy were collected from HNSCC collection in The Cancer Imaging Archive. The cases were split into a training (80%) and a test (20%) dataset. The numbers of recurrence and non-recurrence were 39 (training: 29, test: 10) and 92 (training: 73, test: 19), respectively. Recurrence versus non-recurrence cases were imbalanced. Therefore, the recurrence group in the training dataset was balanced via the synthetic minority over-sampling technique method. A total of 327 dosiomics features enrolled cold spot volume, histogram features, and texture features (gray-level co-occurrence matrix, gray-level run length matrix, gray-level size zone matrix, neighborhood gray tone matrix) were extracted from the original dose distribution (ODD) and LBP on gross tumor volume, clinical target volume, and planning target volume. The Cox-net algorithm was employed for feature selection and dosiomics signature building in training dataset. The ODD and LBP models using support vector machine for the recurrence prediction were constructed with dosiomics features extracted from ODD and LBP. The prediction performances were evaluated in training and test datasets by the accuracy and the area under the receiver operating characteristic curve (AUC).
Results: The accuracy and AUC in the training dataset were 92% and 0.95, respectively, for the LBP model and 77% and 0.81, respectively, for the ODD model. In the test dataset, the LBP model demonstrated the higher performance (accuracy: 79%, AUC: 0.72) than the ODD model (accuracy: 48%, AUC: 0.71).
Conclusion: The LBP-based dosiomics (LBP model) could be feasible for accurately prediction of recurrence after radiation therapy in HNSCC patients.
Image Processing, Texture Analysis
IM/TH- Image Analysis (Single Modality or Multi-Modality): Imaging biomarkers and radiomics