A Novel Imaging-Genomic Approach to Predict Outcomes of Radiation Therapy
Open AccessIntroduction: Tumor regions are populated by various cellular species. Intra-tumor radiogenomic heterogeneity can be attributed to factors including variations in the blood flow to the different parts of the tumor and variations in the gene mutation frequencies. This heterogeneity is further propagated by cancer cells which adopt an “evolutionarily enlightened” growth approach. This growth, which focuses on developing an adaptive mechanism to progressively develop a strong resistance to therapy, follows a unique pattern in each patient. This makes the development of a uniform treatment technique very challenging and makes the concept of “precision medicine”, which is developed using information unique to each patient, very crucial to the development of effective cancer treatment methods. This study aims to determine, for head-and-neck squamous cell carcinoma (HNSCC), whether information present in the heterogeneity of tumor regions in the pre-treatment PET scans of patients and in their gene mutation status can predict the efficacy of radiation therapy in their treatment. Patients found to be unlikely to benefit from radiation therapy would thus be spared unnecessary radiation dose. Materials and methods: Our radiomics analysis was developed using PET scans for 20 patients from the HNSCC database from TCIA (The Cancer Imaging Archive). Clinical data were used to divide the patients into two categories based on the recurrence status of the tumor. Radiation structures are overlain on the PET scans for tumor delineation. Texture features extracted from tumor regions are reduced using correlation matrix-based technique and are classified by methods including Weighted KNN, Linear SVM and Bagged Trees. Slice-wise classification results are computed, treating each slice as a 2D image and treating the collection of slices as a 3D volume. Patient-wise results are computed by a voting scheme which assigns to each patient the class label possessed by more than half of its slices. After the voting is complete, the assigned labels are compared to the actual labels to compute the patient-wise classification accuracies. This workflow was tested on a group of 53 patients of another database: Head-Neck-PET-CT. We then developed a radiogenomic workflow by combining gene expression features with tumor texture features for a group of 11 patients of our third database which is common between TCIA and The Cancer Genome Atlas (TCGA). This database is TCGA-HNSC. We developed a geometric transform-based database augmentation method and applied it to generate additional PET scans using images from the existing dataset. To evaluate our analysis, we tested our workflow on cancer patients at several clinical locations, and included several types of cancer and two imaging modalities. We included PET scans for 24 lung cancer patients (15 from TCGA-LUSC (Lung Squamous Cell Carcinoma) and 9 from TCGA-LUAD (Lung Adenocarcinoma) databases). We used wavelet features along with the existing group of texture features to improve the classification scores. Further, we used non-rigid transform-based techniques for database augmentation. We also included MRI scans for 54 cervical cancer patients (from TCGA-CESC [Cervical Squamous Cell Carcinoma and Endocervical Carcinoma] database) in our study and employed Fisher-based selection technique for reduction of the high-dimensional feature space. Results: The classification accuracy obtained by the 2D and 3D texture analysis is about 70% for slice-wise classification and 80% for patient-wise classification for the head and neck cancer patients (HNSCC and Head-Neck-PT-CT databases). The overall classification accuracies obtained from the transformed tumor slices are comparable to those of the original tumor slices. Geometric transformation appears to be an effective method for database augmentation. The addition of binary genomic features to the texture features (TCGA-HNSC patients) increases the classification accuracies (from 80%-100% for 2D and from 60%-100% for 3D patient-wise classification). The classification accuracies increase from 58% to 84% (2D slice-wise) and from 58% to 70% (2D patient-wise) in the case of lung cancer patients with the inclusion of wavelet features to the existing texture feature group and by augmenting the database (non-rigid transformation) to include equal number of patients and slices in the recurrent and non-recurrent categories. The accuracies are about 64% for 2D slice-wise and patient-wise classification for cervical cancer patients (using correlation-matrix based feature selection) and increase to about 72% using Fisher-based selection criteriaConclusion: Our study has introduced the novel approach of fusing the information present in The Cancer Imaging Archive (TCIA) and The Cancer Genome Atlas (TCGA) to develop a combined imaging phenotype and genotype expression for therapy personalization. Texture measures provide a measure of tumor heterogeneity, which can be used to predict recurrence status. Information from gene expression patterns of the patients, when combined with texture measures, provides a unique radiogenomic feature which substantially improves therapy response prediction scores.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.