Stereotactic radiosurgery cures brain metastases in 80% of patients, but when a treated lesion later enlarges on a scan, doctors cannot tell whether the cancer has returned or harmless radiation necrosis has set in. This uncertainty forces patients into months of repeat scans, extra hospital visits, and sometimes unnecessary repeat radiosurgery—costly interventions that cause anxiety and waste NHS resources. The problem is growing as better cancer treatments keep more patients alive long enough to develop brain metastases. The research team will analyse MRI scans, clinical records, and treatment details from roughly 500 NHS patients whose lesions enlarged after radiosurgery. Three radiologists will assess visual features, a machine-learning pipeline will extract computer-generated radiomic signatures, and statistical models will combine these with patient and treatment factors. The goal is a single, upfront prediction of whether the enlargement is recurrence or necrosis. If the model works, patients with radiation necrosis would avoid further scans and procedures, while those with true recurrence would receive timely repeat treatment. The NHS would save the costs of unnecessary monitoring and interventions. The model would be built entirely from routine care data, making it immediately deployable in existing NHS imaging systems.
View original technical description
Background: With effective cancer treatments, the number of patients living with brain metastases has risen considerably. Stereotactic radiosurgery (SRS) targets cancer deposits precisely and is often (80%) curative. Even with optimally targeted SRS a complication may occur, manifesting as radiation necrosis (RN). RN is a complex side effect, because it can mimic tumour recurrence on imaging. Aims and objectives: The aim of this research is to develop a model for the distinction of RN from recurrent cancer based on MRI scans, alone or in combination with further data. We will address the following research objectives: I. Which MRI parameter(s) is predictive of the correct diagnosis (RN or recurrence)? a) using reproducible visual parameters and DWI metrics b) using computer generated parameters (radiomics/AI) II. Which clinical factor(s) is predictive of the correct diagnosis? III. Which SRS parameter(s) is predictive of the correct diagnosis? IV. Using single or multiple (Ia/b+II+III) parameters, which statistical model performs best? Methods: This study represents an analysis of data acquired in NHS routine care with ethics approval already granted in an open study. Data collection: A dedicated digital imaging platform (https://robin.nottingham.ac.uk/xnat/) is active with research software installed, which permits all measurements in this study. Imaging at the timepoint of treated lesion enlargement, and within 3 months thereafter will be analysed by 3 radiologists blinded to treatment/clinical data. From up to 24 feature categories, those with substantial observer agreement (defined as Cohen s kappa >0.6, intraclass correlation coefficient (ICC)>0.8) will be selected for modelling. A machine learning pipeline will be developed for radiomics. This will utilise a brain tumour segmentation method pre-trained on the brain metastasis dataset of the BraTS-METS Challenge 2023, followed by feature extraction using the PyRadiomics toolkit. Risk factors for the development of RN based on SRS details (e.g. number of treatments, dose details, fractionation) and clinical factors (e.g. systemic therapies) will be documented, informed by literature. Models: Candidate factors will enter univariable analysis. If significant (p<0.05), variables will proceed to multivariable binomial logistic regression to predict the correct outcome (RN or recurrence). Data scoping identified ~500 cases, which substantially exceeds the number of variables (n~20) expected to become eligible for modelling. A proportion (80%) of the cohort will be subjected to statistical analysis to appraise and refine the logistic model before testing on a holdout sample (20%) of the dataset. The final model will be chosen by backwards elimination applying goodness of fit measures. Reference standard The reference standard will be established against strict consensus criteria, by analysing serial MRI evolution and tissue results. If a tissue result is available, this will be prioritised as the reference standard. Anticipated benefit and dissemination: An up-front prediction will deliver 3 key improvements to the patient pathway: Emotional well-being through certainty of the diagnosis, particularly for non-tumour Avoidance of excess MRI, MDT and consultation costs arising from intensified monitoring Ensuring that costly SRS repetition is reserved for definite cancer recurrence The NIHR Impact toolkit will inform the outputs of this research.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know