Performance of machine learning-augmented analysis of radiomics for the head and neck cancer histopathological diagnosis: A systematic review and meta-analysis.

e18526 Background: The imaging criteria used for head and neck cancers (HNC) staging are mostly anatomical with basic quantitative measures, such as size, and admittedly radiologists’ reading of images is dependent on their expertise level. Radiomics, a term referring to extracting and investigating higher dimensional data from images, has been suggested to address these shortcomings. Assisted by machine learning (ML), highly efficient prediction models could revolutionise our diagnostic practices. Our goal was to study the role of ML in the histopathological diagnosis of HNC based on radiomics. Methods: A systematic review and meta-analysis was conducted using electronic databases (PubMed, Scopus, EMBASE, Google Scholar) and including MRI, PET, and CT studies in patients with HNC. Our study was aimed only at diagnosis utilising radiomics and artificial intelligence (ML). A PRISMA diagram retracing the steps of this search process was completed. QUADAS-2 and EQUATOR checklists were completed. A weighted mean, a mean and a median of the performance indicators were recorded. Results: 7 studies were found eligible for meta-analysis. Patient sample sizes ranged between 2-107 patients (median: 18). CT was the most common modality used (4/7 studies). All but one studies were retrospective. Support vector machine and random forest techniques were the main ML techniques used but how the model was built was rarely described. Furthermore, studies did not make clear the exact number of patients in the testing set. Other issues included the reporting of the final model performance with few studies reporting confidence intervals and 2 studies not reporting the exact performance metrics. The accuracy values for the testing set ranged from 58% -94.1%. The meta-analysis showed an overall weighted-mean accuracy of 78.53%, a mean of 82.9% and a median of 84.4%. The weighted mean of the sensitivity was 76.5%, the mean was 83.3%, and for specificity was 83.9% and 88.5%., respectively. The AUC was 0.8. The neuroradiologists’ overall accuracy was 50.4% if weighted, and 54.5% if not, and the corresponding accuracy of the ML classifiers were 78.4% and 79.6%. The ML scored an accuracy of 20% higher than the radiologists. Conclusions: The results are overall encouraging, keeping in perspective the possible calculation biases and small number of studies. There is need for better documentation and standardisation of the applied ML models, which show initially superior performance compared to radiologists.