The potential use of deep learning in performing autocorrection of setup errors in patients receiving radiotherapy

Modern radiotherapy practice relies on multiple approaches for verification of patient positioning. All of these techniques require experienced radiotherapists who understand the anatomical landmarks and the limitations of the used verification techniques. We explore the feasibility of using Artificial intelligence in assisted patient positions using acquired port images (PFIs) and digital reconstructed radiographs (DRRs).

A retrospective study was conducted on patients with brain and aerodigestive tract malignancy who were treated with radiotherapy between 2018 and 2023. A neural network was built to examine and perform auto-correction of the misaligned PFIs and DRRs images. The performance of the neural network was assessed quantitatively by mean-absolute errors (MAE) and mean-squared errors (MSE), and qualitatively by a survey which was sent to 30 experienced medical professionals in the field of radiation therapy.

The total number of patients included in this study was 156 patients. 96 of the patients were treated for aerodigestive tract malignancy while the remaining were treated for brain tumours. The neural network achieved MAE of 27.430 and 27.437 for training and validation sets, respectively, and MSE of 0.5505, and 0.5565 for training and validation sets, respectively. Nineteen medical professionals responded to the survey. They reported a median accuracy score of 8 out of 10.

Our neural network is just one step further in the automation of modern radiotherapy services by using AI-assisted correction of setup errors.

This study demonstrated the potential role of AI in assisting radiotherapists with patient positioning corrections during radiotherapy treatment. Further research is needed to validate the effectiveness of this approach in clinical practice.