Radiomics for the differential diagnosis of radio-necrosis vs. progression after stereotactic radiotherapy of oligometastasis

Radiotherapy is central in the management of many diseases of the brain. In the context of brain tumors, it strongly relies on neuroimaging for diagnosis, treatment planning and to assess the efficacy of therapies. Radiotherapy can greatly benefit from the analytical advances provided by Artificial Intelligence, with applications ranging from tumor segmentation to classification and treatment effect prediction. In the present study, we are addressing the unmet challenge of differentiating tumor recurrence from radiation-induced necrosis after RadioTherapy (RT). Without the invasive pathological confirmation (gold standard), decision-making in this context, based on radiology and clinical examination alone, is often inappropriate and delayed in case of image modification of targeted lesions. We have thus proposed radiomics, a method used to systematically extract hidden features from radiological images, to establish signatures in postop MRI to support the differential diagnosis after RT. In an initial retrospective study, we achieved promising performance of the models trained. These results are now calling for confirmation at a larger scale and through a prospective clinical trial.
We thus propose to improve our existing models by exposing them to larger cohorts, richer data sources and state-of-the-art architectures, preparing them for clinical use by addressing accuracy, interpretability and resilience, and by validating them in a prospective clinical trial.
Improving the performance of our models can serve to greatly improve outcome for the patients. Ultimately, we wish to complement radiologist interpretation with radiomics-based predictions to enrich the diagnostic/prognostic performance, especially when targeted lesions or anatomic areas change. Early management of toxicity or tumor relapse offers the best chances of local control and quality of life.