| BackgroudFocal cortical dysplasia(FCD)is the main cause of refractory epilepsy.MRI plays an important role in preoperative evaluation.However,presurgical identification of MRI abnormalities in FCD type Ⅲ remains difficult.Analyzing the distinctions between these subtypes may provide new insights of clinical,imaging and outcome in patients with FCD type Ⅲ.A sizable number of patients with refractory epilepsy continue to have postsurgical seizures.Deep learning(DL)models can automatically assess complex medical image characteristics and predict prognosis with higher efficiency.ObjectiveThe study was to explore the relationship between preoperative MRI manifestations,pathology of focal cortical dysplasia(FCD)type Ⅲ and the prognosis of postoperative epilepsy.In addition to determine the feasibility of using a DL approach to predict clinically recurrence of epilepsy from T2WI FLAIR in patients with FCD type Ⅲ.Methods266 patients with medically refractory epilepsy were studied retrospectively about the clinical characteristics and MR findings with FCD type Ⅲ diagnosed on pathological examination at least 1 years of post-surgery.MRI images were analysis retrospectively.A convolutional neural network was used for classification of lesions on T2WI FLAIR.The pre-processed original image and the outlined ROI by clinicians were input into our neural network respectively.In this way,the brain regions can be given greater weight,and assist the network to judge whether a subject can achieve seizure freedom after surgery.The models’ performance was analyzed in terms of accuracy,sensitivity,specificity,receiver operating characteristic(ROC)curves and areas under the receiver operating characteristic curve(AUCs).ResultsA total of 266 patients of FCD type Ⅲ included in this study and the median follow-up time was 30 months(range,12-48 months).Age at onset ranged from 1.0 years to 58.0 years,with a median age of 12.5 years.The number of patients under 12 years old was 133(50%)in patients with FCD type Ⅲ.A history of febrile seizures was present in 42(15.8%)cases.In the entire postoperative period,179(67.3%)patients were seizure-free.Factors with P<0.15 in univariate analysis,such as age of onset of epilepsy(P=0.145),duration of epilepsy(P=0.004),febrile seizures(P=0.150),MRI negative(P=0.056),seizure type(P=0.145)and incomplete resection were included in multivariate analysis.Multivariate analyses revealed that MRI negative finding of FCD(OR 0.34,95%CI 0.45-0.81,P=0.015)and incomplete resection(OR 0.12,95%CI 0.05-0.29,P<0.001)are independent predictors of unfavorable seizure outcomes.The overall performance has met these metrics:an area under the receiver operating characteristic curve(AUC)of 96.22%,a sensitivity of 84.47%,and a specificity of 97.21%of the original image only inputs,an AUC of 94.76%,a sensitivity of 84.92%,and a specificity of 96.24%of ROI Only inputs,and AUC of 97.71%,a sensitivity of 90.86%,and a specificity of 96.63%of combined inputs.The performance of different networks with combined inputs.Our proposed network obtained a better result than Densenet and Resnet.Conclusion:MRI-negative finding of FCD lesions and incomplete resection were the most important predictive factors for poor postoperative seizure outcome in patients with FCD type Ⅲ.Deep learning used with conventional MRI can predict the recurrence condition of epilepsy effectively.Artificial intelligence may help the design of clinical management and enable more precise and individualized prediction for postsurgical prognosis of FCD type Ⅲ-related refractory epilepsy. |
PDF Full Text Request