Pretrained Resnet for surgical Outcome prediction using PET Hypometabolism and Excised Tissue (PROPHET)

This is a prospective cohort study of patients who are undergoing resective epilepsy surgery for drug resistant epilepsy, designed to prospectively validate the clinical tool PROPHET, which predicts outcome following epilepsy surgery. PROPHET is a deep-learning based tool that utilises the preoperative T1-weighted MRI, preoperative 18F-FDG-PET scan, and a mask of the resection region as inputs to develop a predicted classification of the outcome following epilepsy surgery (Engel 1 vs Engel 2-4). PROPHET was developed by our team using retrospective data, and in model development, the mask of the resection region was generated from the difference map between the preoperative and postoperative T1-weighted MRIs, and therefore represented the actual resection region. However, to be clinically useful prior to epilepsy surgery, this tool will use a mask of the intended resection region in place of the actual resection region. For all participants in this study, the intended resection region will be annotated manually prior to epilepsy surgery using brain imaging software such as ITK-SNAP or BrainLab. A prediction of the surgical outcome will be subsequently generated using the annotated intended resection region, preoperative MRI and preoperative PET (Model A). The predicted outcome generated by PROPHET will not be re-identified with study participants. Therefore, it will not be disclosed to the treating epileptologist or neurosurgeon prior to, or following, epilepsy surgery. The predicted outcome generated by PROPHET will not impact the clinical decision making about whether the patient will proceed to epilepsy surgery or the nature operation to be performed. The PROPHET predicted outcome will also not be shared with the participant. We will subsequently observe the actual surgical outcome following epilepsy surgery up to 12 months, as measured by the Engel Surgical Outcome Scale. We will also collect the postoperative MRI in the subset of subjects for whom a postoperative MRI was acquired for clinical purposes. The actual resection region will be generated for all patients who have a postoperative MRI, by generating a difference map between the preoperative and postoperative MRI (using the same method as in PROPHET development). Model B will be a repeat prediction using the preoperative MRI, preoperative PET and actual resection region. For both models (model A using the intended resection region, and model B the actual resection region), the predicted Engel outcome and the actual Engel outcomes will be compared using area under the receiver operator characteristic (AUC) to validate the model.