A deep learning framework for ¹⁸F-FDG PET imaging diagnosis in pediatric patients with temporal lobe epilepsy

We retrospectively reviewed a dataset of 201 pediatric patients with TLE (92 girls and 109 boys, age 11.19 ± 3.43 years) who underwent ¹⁸F-FDG PET-CT, EEG, and MRI from November 2013 to April 2019. Exclusion criteria were the following: (1) poor image quality, e.g., severe image artifacts due to head movement, and (2) negative finding on ¹⁸F-FDG PET scan, (3) incomplete clinical data, and (4) missing EEG or MRI data. To determine a negative finding (no significant abnormal finding) on ¹⁸F-FDG PET scan, all images were reviewed by the team of experts to reach a consensus. A total of 136 cases were included for the following analysis (including 57 girls and 79 boys with ages of 10.84 ± 3.35 years).

An age-matched control group dataset (6 girls and 18 boys, age 11.58 ± 2.83 years) was built by retrospectively reviewing patients with extracranial lymphoma who had no history of neurologic disorders, psychiatric illnesses, chemotherapy, or radiotherapy. This retrospective study was approved by the Human Subject Research Ethics Committee of The Second Hospital of Zhejiang University School of Medicine, and the requirement of informed consents was waived (Approval Number: 2019162).

All subjects were injected with 3.7 MBq/Kg ¹⁸F-FDG after fasting at least 6 h and then rested with eyes closed in dark and quiet environment for a 40-min uptake period. PET-CT brain images were acquired on a PET-CT scanner (Biograph mCT; Siemens Medical Solutions), using a 5-min bed position and 3D whole-head acquisition. Data were reconstructed by a vendor-provided software using Fourier recombination (FORE) and attenuation-weighted ordered-subsets expectation maximization (AW-OSEM). The reconstructed resolution of PET images was almost isotropic with a full width at half maximum (FWHM) of 4.4 mm in the center and a FWHM of 4.8 mm at 10 cm off-axis. Images were then resliced into 2-mm-thick slices and reoriented to the coronal-transaxial-sagittal orientation on Siemens medical workstation.

¹⁸F-FDG PET images were visually assessed by two physicians specialized in epilepsy diagnosis (YL, XW) who were unblinded to the seizure semiology, EEG, and MRI findings. All epileptic foci were manually delineated using ITK-SNAP software (version 3.6.0; https://​www.​itksnap.​org). For inconsistent cases, the final decision was made by the team of experts to reach a consensus through cross-review.

The ¹⁸F-FDG PET images were spatially normalized to an in-house symmetric template by the symmetric diffeomorphic registration using Advanced Normalization Tools (version 2.1; https://​stnava.​github.​io/​ANTs) [21]. All warped images were then transformed to Z-maps using Fisher’s Z-transformation. After that, PET images were randomly divided into two sets: training (106 patients and 9 controls) or testing (30 patients and 15 controls) set. The training set was involved in the model training for both radiomics analysis and deep learning, while the testing set was used to evaluate the performance of well-trained models.

The radiomics-based ¹⁸F-FDG PET analysis consisted of feature extraction, feature selection, and model training. In the feature extraction, the symmetricity features of the bilateral temporal lobes were quantitatively investigated to imitate the visual assessment that bilateral glucose uptake was compared simultaneously. The radiomics features (totally 386 dimensions, described in Supplementary Materials), including 344 multiscale wavelet features, 36 texture features, and 6 intensity features, were extracted from the bilateral temporal lobes. The multiscale wavelet features included intensities and textures from Coiflet 1 wavelet (coif1) transformed images. The texture features were calculated based on histogram, gray-level co-occurrence matrix (GLCM), neighborhood gray-tone difference matrix (NGTDM), and gray-level zone size matrix (GLZSM). The intensity features included mean, variance, skewness, kurtosis, energy, and entropy. The symmetricity feature was obtained by calculating the absolute distance between the radiomics features of the left and right lobes according to the following equation:where \( {f}_i^{\mathrm{left}} \) and \( {f}_i^{\mathrm{right}} \) represent the number i feature of either the left or right lobe. The symmetricity vector d = (d₁, d₂, ⋯, d₃₈₆) was then built to represent the individual inter-hemispheric temporal lobe symmetricity and formed a symmetricity distance matrix. Column-wise Z-transformation was applied to the symmetricity distance matrix to obtain the normalized symmetricity matrix.

For feature selection, the hierarchical clustering (HC)—a widely used clustering method—was applied to the normalized symmetricity matrix to obtain k clusters (krepresents the number of clusters). In each cluster, the most significant symmetricity feature was selected by maximum absolute Fisher inter-intra class variance ratio (FICVR) [22], which can be defined in the following equation:

where μ_p stands for the mean symmetricity features of the patient and μ_c represents that of the control, while σ_p and σ_c denote the standard deviations of the patient and the control, respectively. The FICVR of a feature represents its capability of distinguishing the patient from the control.

To verify the advantages of our selected features, this study evaluated the classification performances of our selected features, low-order features (intensity features), mono-features (features with favorable classification metrics), and multi-features (all features and principal component analysis (PCA) selected features) based on logistic regression models. The classification metrics such as accuracy, sensitivity, specificity, and receiver operating characteristic (ROC) were displayed. To improve the robustness of the logistic regression models, the training set was divided into 10 groups through a 10-fold cross-validation procedure, which was consisted of 10 iterations: in each iteration, one group was used as the validation set, while the rest of the training sets were used for model training [23]. This procedure was repeated for 10 times to obtain the average performance.

A deep learning approach, symmetricity-driven Siamese CNN, was proposed to localize the epileptic focus. This proposed framework, which consisted of two identical 18-layer residual convolution neural networks (ResNet) [24], was firstly used to extract deep features of bilateral image cubes automatically and then to localize the focus by calculating the feature differences (Fig. 1). The procedures of training and inference were demonstrated in the red and black dashed-dotted boxes, respectively. In the procedure of training, the input x belong to the sample image set X; while the corresponding outpu y belong to the label set Y. It is noted that y = 0.0 and y = 1.0 represent the normal and abnormal image cubes, respectively. Therefore, the epileptic focus detection could be considered as a symmetricity-driven binary classification. In the procedure of inference, the output probability p of the framework was obtained on testing set. The above-mentioned procedures were implemented by using PyTorch (version 1.0; https://​pytorch.​org).

To obtain the inputs of Siamese CNN, the right and left parts of PET images were partitioned to pairs-of-cubes (PoCs), and each PoC was consisted of n × n × n pixels (here we set n = 48). Then the positive (with focus) and negative (without focus) PoCs were obtained from the patients and controls. To balance the sample sizes of positive and negative PoCs, data augmentation was conducted on the training set, such as flipping, radial distortion, and intensity modification. In addition, sample weighting was used by setting relatively larger weights for minority samples in each training batch. The PoC-based training strategy could overcome the challenge of small training sample size in PET imaging and improve the framework robustness.

In the process of Siamese CNN training, PoCs were respectively fed to the two sub-networks to obtain two 1024-entry feature vectors. The absolute difference between the two feature vectors was then processed by a fully connected layer to get the prediction score. For parameters setting, we applied the standard stochastic gradient descent with learning rate of 0.01 and set weight decay as 0.005, batch size as 128, and training epoch as 16. In each training epoch, the trained model obtained the validation results on the validation set. The model with the best validation result was selected as the final model for the inference on the testing set.

In the procedure of inference for focus detection, the input PET images were partitioned into cubes to generate PoCs. The aforementioned procedure was executed on each PoC to obtain the output p. Eventually, a probability heat map for predicting the abnormal focus was achieved in one PET image, as illustrated in the last stage of Fig. 1. To evaluate the performance of our proposed method on PoC classification, this study calculated the Jensen-Shannon (JS) divergence of PoCs and applied a logistic regression (LR) classifier to discriminate normality and abnormality for comparison.

After detecting abnormal focus, the framework further calculated the SUVR and determined hyper- or hypometabolism by computing asymmetric index (AI):

If absolute AI was larger than the threshold (e.g., 0.15) for three consecutive slices, the focus was determined as severe metabolic abnormality, otherwise mild. The hypo- or hypermetabolism was determined by the sign of AI, i.e., negative for hypo- and positive for hyper.

To evaluate the accuracy and consistency of the proposed method, results were compared with those of SPM (version 8; http://​www.​fil.​ion.​ucl.​ac.​uk/​spm/​) and physicians with different experience levels, i.e., physicians specialized in epilepsy diagnosis (specialist), junior nuclear medicine physicians (junior), senior nuclear medicine physicians (senior), and a senior neurologist (neurologist). In SPM analysis, all PET images were spatially normalized to standard stereotactic space by an in-house symmetric ¹⁸F-FDG PET template, followed by smoothing with 8-mm FWHM Gaussian kernel. Two-tailed inference and cluster size level of 100 were used as prior study [25]. SPM results were classified into four levels according to the uncorrected P values, i.e., normal (P > 0.05), probably normal (0.01 < P ≤ 0.05), probably abnormal (0.001 ≤ P < 0.01), and abnormal (P < 0.001).

The diagnosis and localization results of proposed method were compared with those of SPM analysis and physicians with different levels. The physicians decided each case as normal, probably normal, probably abnormal, or abnormal and then manually delineated the epileptic foci for all suspected “abnormal” cases. Dice coefficient [26] was used to quantify the similarity of proposed method, physicians, and SPM results to the reference standard. Wilcoxon signed-rank test was employed to compare the dice coefficients of different methods. The percentage of correct diagnosis, ROC curves, sensitivity, and specificity were reported.

The glucose metabolic levels were classified into four categories: severe hypometabolism, mild hypometabolism, mild hypermetabolism, or severe hypermetabolism. The metabolic abnormality level determination results were compared between proposed method and physicians. Physicians first found the epileptic foci and estimated the metabolic abnormality levels of the foci blinded to clinical information including seizure semiology, EEG, and MRI findings. After that, physicians utilized the clinical information to re-evaluate foci and metabolic levels. At last, we compared the physician diagnosis results blinded or unblinded to the clinical information and those of the proposed framework. Wilcoxon signed-rank test was employed to evaluate physician determination of hypo- and hypermetabolism blinded or unblinded to clinical information. McNemar’s test was employed to compare physician visual assessments and proposed method. P value less than 0.05 (P < 0.05) was considered statistically significant.

The heat map of symmetricity features was demonstrated in Fig. 2a. The rows represented images of the control and the patient, and columns represented symmetricity features. The colors of the heat map varied from orange to purple, which represented the values of symmetricity features changed from small to large. All symmetricity features were grouped into 10 clusters by HC, from which the feature with maximum FICVR was selected (totally 10 features were selected, see Table S1 in Supplementary Materials). As shown in Fig. 2b, the ROC curve of our proposed method based on 10 selected features outperformed those classification methods using 4 single features in mono-feature category with largest area under curves (AUCs). Table 1 illustrated all classification metrics of low-order, mono-, multi-, and our selected features. The “high gray-level zone emphasis” feature had the highest sensitivity (0.82); the proposed classification method had the highest AUC (0.92), accuracy (0.81), and specificity (0.89), which suggested the proposed cluster-based feature selection method could purposely select the most TLE-correlated symmetricity features.

The diagnosis results of the proposed method, the SPM analysis, and visual assessments were shown in Table 2 and Fig. 2c. The experienced physicians had relative higher AUC and accuracy than the ones with less experience in epilepsy diagnosis. The SPM analysis had similar AUC and accuracy to the specialists (AUC 0.76 vs. 0.80; Accuracy 0.73 vs. 0.75). The proposed radiomics-based analysis had the highest AUC (0.89), accuracy (0.82), sensitivity (0.84), and specificity (0.80) than the SPM analysis and visual assessments, suggesting a strong correlation between high-dimensional symmetricity features and epilepsy. This finding supported our hypothesis that epilepsy had intrinsic high-dimensional metabolic asymmetricities that cannot be easily recognized by visual assessment.

A total of 51,689 PoCs were extracted from 106 patients and 9 controls, in which 60% of PoCs were used for training, 20% of PoCs for validation, and 20% for testing. The proposed method could accurately detect abnormal PoCs (AUC = 0.93), which is better than the JS-LR (AUC = 0.72) (Fig. 3a). As shown in Fig. 3b, the results demonstrated that the proposed approach had the highest average dice coefficient (0.51) compared with the physicians (0.31–0.44) and SPM analysis (0.24). Wilcoxon signed-rank test showed that proposed method significantly outperformed SPM analysis (P < 0.01), junior physicians (P = 0.005–0.017), and neurologist (P = 0.038). The proposed method had a marginally significantly higher dice coefficient than senior physician (P = 0.068) and specialists (P = 0.09–0.17). Two examples of hypo- and hypermetabolism were presented in Fig. 4. Both hypo- and hypermetabolism regions predicted by the proposed deep learning method were consistent with the reference label. Moreover, the proposed method tended to be more aggressive by including slightly larger regions than the reference label to avoid false negatives (i.e., missed diagnosis). Such aggressiveness can be fine-tuned by a threshold parameter used in the framework.

The determination ability of metabolic abnormality level was compared between the proposed method and the physicians (see Fig. 5). All physicians failed to recognize mild hypermetabolism with or without access to the clinical information. All physicians (except for one junior nuclear medicine physician) had higher determination accuracy in terms of severe hypometabolism than in the mild ones (71% vs. 55%; P = 0.032). Consulting clinical information can significantly increase the determination accuracy of physicians for mild hypometabolism from 46 to 64% (P < 0.01). However, the impact of clinical information for severe hypermetabolism determination is very limited (Accuracy = 78% for both blinded and unblinded).

The proposed method had higher determination accuracy for both severe and mild abnormalities (94% and 85%, respectively) than physicians blinded (68% and 42%) to the clinical information or unblinded (75% and 59%), as presented in Table 3. The McNemar’s test showed that the proposed method significantly outperformed the physicians blinded or unblinded to clinical information (90% vs. 56% and 68%; P < 0.01). In particular, the proposed method successfully determined all the hypermetabolism (100%), which was superior to all the physicians.