Understanding the Impact of Statistical and Machine Learning Choices on Predictive Models for Radiotherapy

Abstract

During radiotherapy (RT) planning, an accurate description of the location and shape of the pelvic organs is a critical factor for the successful treatment of the patient. Yet, during treatment, the pelvis anatomy may differ significantly from the planning phase. A series of recent publications, such as PREVIS [FMCM ∗ 21], have examined alternative approaches to analyzing and predicting pelvic organ variability of individual patients. These approaches are based on a combination of several statistical and machine learning methods, which have not been thoroughly and quantitatively evaluated within the scope of pelvic anatomical variability. Several of their design decisions could have an impact on the outcome of the predictive model. The goal of this work is to assess the impact of alternative choices, focusing mainly on the two key-aspects of shape description and clustering, to generate better predictions for new patients. The results of our assessment indicate that resolution-based descriptors provide more accurate and reliable organ representations than state-of-the-art approaches, while different clustering settings (distance metric and linkage) yield only slightly different clusters. Different clustering methods are able to provide comparable results, although when more shape variability is considered their results start to deviate. These results are valuable for understanding the impact of statistical and machine learning choices on the outcomes of predictive models for anatomical variability.