TY - JOUR T1 - Prediction of human papillomavirus associated oropharyngeal cancer using multiple machine learning algorithms and PET/CT image radiomics features. JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 1185 LP - 1185 VL - 62 IS - supplement 1 AU - Atlas Haddadi Avval AU - Ghasem Hajianfar AU - Mehdi Amini AU - Mehrdad Oveisi AU - Isaac Shiri AU - Habib Zaidi Y1 - 2021/05/01 UR - http://jnm.snmjournals.org/content/62/supplement_1/1185.abstract N2 - 1185Objectives: Since Oropharyngeal cancer (OPC) comes in multiple subtypes, physicians should be cautious when planning the treatment and management of patients. Human Papillomavirus (HPV) positive OPC should be distinguished from HPV negative subtypes. In this regard, we aimed to evaluate the capability of combining multiple feature selection (FS) and machine learning (ML) classifier algorithms for OPC classification (HPV status prediction) based on the extracted radiomics features of CT, PET, and fused PET/CT images. Methods: This retrospective study consisted of 95 patients with OPC all of whom were biopsy-tested for the determination of HPV status. Data were randomly split into training (70%) and testing (30%) sets. Pre-treatment CT and PET images of tumors were collected and fused with Gradient Transfer Fusion (GTF) and Weighted Least Squares (WLS) methods. After reaching the aforementioned four different image sets (CT, PET, GTF, and WLS), a total of 221 radiomic features were extracted from each set, and two FS methods namely minimum redundancy maximum relevance (MRMR) and recursive feature elimination (RFE) were separately applied to the extracted features. Random Forest (RF), Naive Bayesian (NB), Stacked Learning (SL), and XGBoost (XGB) were utilized as the four ML methods. Totaling 32 models, the training set was employed for our hyperparameter optimization (along with 5-fold cross-validation) and the validation set (with 100 times bootstrapping) was used to assess the performance of models. The area under the receiver operating characteristic curve (AUC) and accuracy were reported for each model. Results: Our findings showed that on each of the imaging sets, special combinations of ML and FS algorithms achieved the best performance for HPV status prediction. For instance, both NB and RF machine learning algorithms in combination with the MRMR feature selector achieved an AUC of 0.80 in the CT imaging set (NB also achieved an accuracy of 0.81). In addition, in the GTF image set, SL and RF algorithms gained AUCs of 0.78 and 0.77, respectively, when combined with the RFE feature selector (accuracy of the SL model was 0.77). Conclusions: This study supports the potential of using machine learning models on PET, CT, and fused PET/CT images. Through different feature selection algorithms, these models can predict whether OPC in a patient is associated with HPV or not. Further studies on larger datasets can prove the capability of image-derived radiomic features for non-invasive characteristic prediction of OPC. ER -