Is the use of machine learning in head and neck cancer radiotherapy supported by clinical trials?

Objective

The employment of artificial intelligence (AI) and machine learning (ML) in cancer management is well documented. The diversity and complexity of models, approaches, endpoints and anatomies makes it difficult to draw pertinent conclusions regarding the usefulness of ML in various applications. Given that clinical trials, particularly randomized ones, are the gold standard when testing new techniques, the current work aims to collate trials that investigated the effectiveness of ML in head and neck cancer (HNC).

Methods

A systematic search of Medline/PubMed and Web of Science databases was performed to identify clinical trials reporting the use of AI / ML in HNC for either of the following clinical aspects: detection / classification, image segmentation and treatment response / dose distribution prediction. Of the 2395 identified studies, 42 met the eligibility criteria.

Results

Most studies confirmed the usefulness of ML in HNC clinical applications through reported parameters: accuracy, area under the curve, specificity, sensitivity, dice coefficient. Deep learning models utilizing multiple layered neural networks were often the choice as they outperform machine learning in predictive model development. However, traditional algorithms (KNN, SVM, logistic regression) often showed comparable results with deep learning. The most commonly used radiomic features were related to image texture irrespective of the model’s application.

Conclusions

Clinical trial results show a supportive role for ML in HNC management, whether diagnostic/staging- or treatment-related. The best performing applications are enhancing deep learning algorithms applied to imaging with clinical data for a more comprehensive and better-integrated approach to personalized treatment.