Artificial Intelligence in CNS Radiation Oncology (AI-RAD)

In the proposed retrospective study, patients treated with Radiotherapy (RT) for Central Nervous System(CNS) tumors will be included. The DMG database maintaining records of patients registered and treated in TMC (Tata Memorial Centre) neuro oncology DMG (Disease Management Group) will be screened to identify the patients eligible for the study. With approximately 500-600 patients with CNS tumors treated with RT in TMC every year, we expect a ceiling of 6000 patients during 2010-2022, which will be the maximum number of patients used for the analysis. The images (CT, MRI, PET) used for RT planning, mid-treatment imaging as part of IGRT (Image-guided radiation therapy) or disease evaluation, and response assessment/ surveillance post-RT will be analyzed. The radiation plans and dose-volume histogram will be obtained from TPS (Treatment Planning System). All the images and radiation-related data will be downloaded from the PACS (Picture Archiving and Communication System) and TPS, applying anonymization filters. Clinical features (patient, disease, treatment-related characteristics, and outcomes) will be extracted by review of electronic medical records. Imaging pre-processing will be done, which will include skull stripping and registration across different modalities (e.g., MRI and CT) or different sequences (e.g., T1C, T2W, ADC) will be done using rigid or deformable algorithms as suited best for the modality. The target volumes, i.e., gross tumor volume (GTV), clinical target volume (CTV), and planning target volume (PTV), and OARs will be individually reviewed by radiation oncologists with modifications applied as appropriate (e.g., exclusion of OARs (Organs at risk) distorted by disease or surgery) and will be used to train the machine learning models for supervised learning. The contours and the images will be resampled to a uniform resolution for different sequences or modalities (e.g., T2W/ ADC/ PET) to match either with the 3D sequence (e.g., FSPGR sequence) or available images with the least slice thickness. Subsequently, normalization techniques (e.g., histogram normalization/ Z-score normalization) will be undertaken within the individual images and across the entire dataset to account for image heterogeneity, including field strength for MRI and different image acquisition parameters. For autosegmentation, both supervised and unsupervised machine learning algorithms will be applied. For the supervised model, the entire database will be split into training and test cohorts for the model and application development, respectively. Since the OARs are uniformly applicable for different histology or tumor sites, autosegmentation training will be applied to the entire dataset. However, given there are variations in target volume delineations (e.g., for circumscribed vs. diffuse tumors, low grade vs. high grade), the training/ testing for TVs will be applicable for individual disease entities. The effectiveness of the automated model will be tested using the dice similarity coefficient between manually segmentation regions and AI-based segments. For outcome prediction (e.g., survival and toxicities), the next step will include feature extraction from images (CT, MRI, PET) corresponding to different TV and OARs and RT dose distribution data converted to volumetric image/ number data (dosiomics), which will consist of first-order (including shape, histogram), second-order or higher-order (e.g., different texture features like GLCM, GLDM, GLSZM, etc.), or deep learning techniques will be employed. Delta-radiomics will include temporal changes in the radiomic features from different time points for the same patient within the entire volume and individual regions. Subsequently, feature reduction and selection techniques like LASSO, recursive feature elimination will be used to shortlist the number of features depending on the sample size. The outputs will be decided based on the modeling defined for specific class problems (e.g., tumor vs. edema, recurrence vs. pseudoprogression, outcomes, tumor region of interest vs. non-tumoral area) as obtained from the clinical information. Any class imbalance will be addressed using methods like random subset sampling or SMOTE analysis for data augmentation of the minority class. Machine learning algorithms like LDA, k-NN, SVM, random forest, AdaBoost, etc., will be applied singularly or in combination as an ensembled classifier to find the model with the best performance. Deep learning classifiers will be used along with feature-based modeling and compared to test the classifier's applicability. Validation techniques like leave-one-out validation, k-fold validation, and split (into training and test cohort) will be used to assess the stability of the machine learning model. Radiomic analysis will be done by data scientist/ study investigators with expertise in data analytics. All segmentations will be done on open source software like ITK snap (itksnap.org) or 3D Slicer (slicer.org). Feature extraction and modeling will be done using opensource software like Python (python.org). As a tertiary objective in the project, we will develop a protocol for anonymized data storage (clinical information, radiation planning and response assessment images, radiation planning data, intra-treatment images like cone beam computed tomography) in a secured image biobank repository with protected cloud space. Also, natural language processing (NLP) algorithms will be developed to train and validate model for extraction, and documentation of clinical variables extracted for the study. With continuous advancements in computational science, available newer analytical techniques and platforms will be applied as appropriate by collaborators from Bhabha Atomic Research Centre, Mumbai, by sharing anonymized data.