In our study, we successfully extracted 1132 radiomic features based on PET images and selected 7 features by both Boruta and Lasso regression to generate ML models for identifying the candidate patients of TLE. The performance of 11 algorithms of model construction was compared, LR model was chosen to be the optimal method with the highest AUC. To optimize the LR model, AIC in a Stepwise Algorithm was used to further select the best three radiomic features including original_gldm_DependenceEntropy, log.sigma.5.0.mm.3D_firstorder_Energy, and wavelet.LHL_glrlm_GrayLevelNonUniformity. Extracted from the original image, original_gldm_DependenceEntropy is a feature based on A Gray Level Dependence Matrix (GLDM) which quantifies gray level dependencies in an image. DependenceEntropy is a measure of the randomness/variability in the gray level dependency defined as the number of connected voxels within a distance that is dependent on the center voxel. log.sigma.5.0.mm.3D_firstorder_Energy extracted from the image processed by Laplacian of Gaussian (log) filter with 5.0mm sigma is a measure of the magnitude of voxel values in this image. After the PET image had been managed with a wavelet filter combined with three dimensions low, high, and low respectively, wavelet.LHL_glrlm_GrayLevelNonUniformity could be extracted and represent the similarity of gray-level intensity values in the image, where a lower GrayLevelNonUniformity value correlates with a greater similarity in intensity values. Eventually, a tuned LR model based on three optimal PET radiomic features showed promising results in an independent training and test cohort, with an AUC of 98.1 and 95.7 for distinguishing TLE patients. Our study shows that PET-based radiomics can be useful biomarkers for identifying TLE patients.
Previous studies have shown that 18F-FDG PET is manifested to be an available noninvasive method to complementally assist the diagnosis and prognosis prediction of epilepsy, help intracranial electrode placement, and it can also decrease the amounts of invasive EEG tests that need to be conducted.24–26 However these studies were also focused on single conventional parameters. In this condition, surgery is the most vital treatment for medication-refractory temporal lobe epilepsy, and 18F-FDG PET has important added value for the surgery decision-making in TLE because of the increased predictive values of MRI and video-EEG monitoring in combination with 18F-FDG PET.
18F-FDG PET seemed especially valuable when MRI findings were negative or not concordant with EEG findings.24,27–29 A meta-analysis summarized the performance of 18F-FDG PET imaging analyzed with the traditional method of visual for the localization of epileptogenic zone in patients with epilepsy, considered EEG or surgical outcomes as the gold standard, 18F-FDG PET demonstrated an overall sensitivity of 0.66 (95% CI: 0.58–0.73) and specificity of 0.71 (95% CI: 0.63–0.78), with an AUC of 0.71.30 Quantification of 18F-FDG PET might be helpful to improve the performance for epilepsy diagnosis and the localization of the epileptogenic zone. KL Wang et al31 compared the concordance rates with the gold standard evaluated by visual assessment(40%), statistical parametric mapping (SPM, 83%), and three-dimensional stereotactic surface projection (3D-SSP,71%). The results show that SPM and 3D-SSP both can improve the detection rate of the epileptic focus compared to visual assessment, which demonstrates the great advantage of quantification
Radiomics is a newly developing medical image analysis method with high-throughput extraction of quantitative features and automated quantification of radiographic phenotypes.32 Although the pathophysiology of epilepsy remains poorly understood, dysfunction of cerebral energy metabolism, neuronal loss, and the reduction of synaptic activity may be involved, and PET is a direct reflection of cerebral energy metabolism and has been reported to partially reflect the reduction of synaptic activity.33 Radiomics features, especially high-order features, capture the spatial variation in PET signal intensity that may reflect the underlying pathophysiology, which may explain our observation.
In the present study, we combined the Boruta algorithm and Lasso regression to select radiomic features, resulting in creating a more generalized and stable classifier that is robust against the idiosyncrasies of the training data. Our study shows great potential that using PET radiomics-based ML models to identify TLE patients because of high AUCs ranging from 0.803 to 0.948 in 11 various traditional machine learning algorithms.34 The logistic regression model, which is an efficient and powerful way to assess independent variable contributions to a binary outcome, was the final choice to distinguish TLE patients. Logistic regression can iteratively calculate the strongest linear combination of variables with the greatest probability of detecting the observed outcome.35 Identifying whether the participants are TLE patients or not is the objective of our study, it belongs to a typical binary classification, which may explain why the LR model is the prime choice. Decision tree modeling is a non-parametric supervised learning algorithm, which is utilized for classification tasks and easy to explain visually. Analyzing our data by decision tree modeling, log.sigma.5.0.mm.3D_firstorder_Mean was the root node for classification, wavelet.LHL_glrlm_GrayLevelNonUniformity and original_gldm_DependenceEntropy, which were also selected by LR models were internal nodes. These important PET radiomics features were bi-validated using various classical ML algorithms (figure S1-2). Compared with decision tree modeling, logistic regression belongs to higher AUC, which was more suitable for distinguishing TLE patients. And an online web application to aid the diagnosis based on FDG PET data was developed according to our LR modeling results (https://wane199.shinyapps.io/TLE_Classification/). It could be conducted at different epilepsy centers, which can further validate the performance of our algorithm.
Our study has several limitations. Firstly, the present investigation constitutes a retrospective analysis conducted solely within a singular healthcare facility, featuring a comparatively modest cohort size. Subsequent investigations encompassing a broader dataset and external verification are essential to facilitate an enhanced evaluation. Second, we only extracted the radiomics feature in the temporal lobe region, other important regions such as the hippocampus and para-hippocampus should be investigated in future studies. Third, image feature selection still needs further optimization by deep learning. Fourth, our study focused on the classification performance of PET imaging independently, so we identified the TLE patients only based on PET imaging, clinical information of each patient was not involved. Further studies including clinical data should be performed.