Machine learning approach for Breast Cancer Radiotherapy degradation prediction using clinical data and photos

Abstract

Breast cancer is the most common cancer worldwide but modern treatment has increased the survival rate at 5 years to 90% or more in high-income countries. Most breast cancers are treated with breast-conserving surgery that can be followed by radiotherapy. However, this exposure to radiation may come with short and long-term side-effects that can have an impact on a patient's quality of life. This work investigates the possibility to predict the degradation and the final aspect of the breast after radiotherapy by exploiting AI models trained on small datasets made up of clinical data or pictures taken from the front and the side of Caucasian patients. Unfortunately, the best models only achieve slightly better than random predictions, and with a lot of instability. It was found that the pathological stage, the biomarkers of the removed tumor, the number of sentinel nodes removed, the number of lymph nodes and the indication that a patient is diabetic and/or a smoker, are the attributes that have the biggest impact on predictions. The best model trained on images over-performed the best clinical data trained model. Fine-tuning convolutional neural networks instead of using them as features extractors yields better performance.