Prediction of treatment response and outcome in locally advanced rectal cancer using radiomics and deep learning: an example case to demonstrate a general purpose deep-learning-based processing pipeline for image classification.

点击翻译按钮获取中文摘要

Over the last decade, rectal cancer has become the number 3 most lethal disease in Europe with a 5-year survival rate of only 68% in Germany. Incidence rates are still increasing and remain above 60%. Even though diagnostic and treatment opportunities have improved in recent years, rectal cancer remains a heterogeneous disease in terms of treatment response and outcome.Thus far, only clinical and magnetic resonance imaging (MRI) based criteria are used for guiding treatment decisions, once the diagnosis has been confirmed. Although MRI has evolved to become the standard diagnostic approach in the local staging of rectal cancer, it does not provide information on intratumor heterogeneity or molecular subtypes. Consequently, novel imaging biomarkers are urgently needed in order to better characterize rectal cancer subtypes, aiming at an improved prediction of treatment response and patient outcome.Texture analysis, radiomics and deep learning strategies are increasingly used to address these challenges and to improve patient care. However, patient cohorts in many radiomics studies were relatively small, studies often lacked a validation cohort and imaging data were obtained within one single institution or different centers with similar MRI scanners, thus not allowing for assessing the generalizability of the trained models. Modern radiomics techniques therefore are increasingly shifted towards recent developments in deep learning.The purpose of this study is to develop a radiomics- and deep learning-based imaging signature of rectal cancer, which is able to decode different tumor phenotypes and to non-invasively assess / predict therapeutic response in correlation to histopathology as well as genomics / clinomics. This should lead to a comprehensive characterization of tumor heterogeneity and tumor biology by imaging criteria, which will then allow for individually tailored treatment strategies in the future. The entire framework will be developed based on available open source methodology and will be made available for future use and research, thus enabling future translation into routine clinical practice.Available methodology from the field of radiomics, artificial intelligence, and computer vision will be applied on a prospectively acquired, well-structured multi-center dataset from the CAO-ARO-AIO-12 study, which is already available and accessible for the purpose of the present study. This dataset includes radiation-planning CT data, pre- and post-treatment multiparametric MRI data, histopathological information as well as clinical and genomic data. A team of four experienced PIs from the fields of oncologic imaging, radiomics, MRI physics, radiation therapy, and informatics will work as an interdisciplinary team and share their knowledge-gain with other members of the priority program consortium in an open-source fashion.