RESUMEN
Structural magnetic resonance imaging (sMRI) is an essential part of the clinical assessment of patients at risk of Alzheimer dementia. One key challenge in sMRI-based computer-aided dementia diagnosis is to localize local pathological regions for discriminative feature learning. Existing solutions predominantly depend on generating saliency maps for pathology localization and handle the localization task independently of the dementia diagnosis task, leading to a complex multi-stage training pipeline that is hard to optimize with weakly-supervised sMRI-level annotations. In this work, we aim to simplify the pathology localization task and construct an end-to-end automatic localization framework (AutoLoc) for Alzheimer's disease diagnosis. To this end, we first present an efficient pathology localization paradigm that directly predicts the coordinate of the most disease-related region in each sMRI slice. Then, we approximate the non-differentiable patch-cropping operation with the bilinear interpolation technique, which eliminates the barrier to gradient backpropagation and thus enables the joint optimization of localization and diagnosis tasks. Extensive experiments on commonly used ADNI and AIBL datasets demonstrate the superiority of our method. Especially, we achieve 93.38% and 81.12% accuracy on Alzheimer's disease classification and mild cognitive impairment conversion prediction tasks, respectively. Several important brain regions, such as rostral hippocampus and globus pallidus, are identified to be highly associated with Alzheimer's disease.
Asunto(s)
Enfermedad de Alzheimer , Disfunción Cognitiva , Humanos , Enfermedad de Alzheimer/diagnóstico por imagen , Imagen por Resonancia Magnética/métodos , Hipocampo , Encéfalo/diagnóstico por imagen , Encéfalo/patología , Diagnóstico por Computador , Disfunción Cognitiva/diagnóstico por imagenRESUMEN
Objective: Focal cortical dysplasia (FCD) is the most common pathological cause for pediatric epilepsy, with frontal lobe epilepsy (FLE) being the most prevalent in the pediatric population. We attempted to utilize radiomic and morphological methods on MRI and PET to detect FCD in children with FLE. Methods: Thirty-seven children with FLE and 20 controls were included in the primary cohort, and a five-fold cross-validation was performed. In addition, we validated the performance in an independent site of 12 FLE children. A two-stage experiments including frontal lobe and subregions were employed to detect the lesion area of FCD, incorporating the asymmetric feature between the left and right hemispheres. Specifically, for the radiomics approach, we used gray matter (GM), white matter (WM), GM and WM, and the gray-white matter boundary regions of interest to extract features. Then, we employed a Multi-Layer Perceptron classifier to achieve FCD lesion localization based on both radiomic and morphological methods. Results: The Multi-Layer Perceptron model based on the asymmetric feature exhibited excellent performance both in the frontal lobe and subregions. In the primary cohort and independent site, the radiomics analysis with GM and WM asymmetric features had the highest sensitivity (89.2 and 91.7%) and AUC (98.9 and 99.3%) in frontal lobe. While in the subregions, the GM asymmetric features had the highest sensitivity (85.6 and 79.7%). Furthermore, relying on the highest sensitivity of GM and WM asymmetric features in frontal lobe, when integrated with the subregions results, our approach exhibited overlaps with GM asymmetric features (55.4 and 52.4%), as well as morphological asymmetric features (54.4 and 53.8%), both in the primary cohort and at the independent site. Significance: This study demonstrates that a two-stage design based on the asymmetry of radiomic and morphological features can improve FCD detection. Specifically, incorporating regions of interest for GM, WM, GM, and WM, and the gray-white matter boundary significantly enhances the localization capabilities for lesion detection within the radiomics approach.
RESUMEN
OBJECTIVE: Precise preoperative evaluation of drug-resistant epilepsy (DRE) requires accurate analysis of invasive stereoelectroencephalography (SEEG). With the tremendous breakthrough of Artificial intelligence (AI), previous studies can help clinical experts to identify pathological activities automatically. However, they still face limitations when applied in real-world clinical DRE scenarios, such as sample imbalance, cross-subject domain shift, and poor interpretability. Our objective is to propose a model that can address the above problems and realizes high-sensitivity SEEG pathological activity detection based on two real clinical datasets. METHODS: Our proposed innovative and effective SEEG-Net introduces a multiscale convolutional neural network (MSCNN) to increase the receptive field of the model, and to learn SEEG multiple frequency domain features, local and global features. Moreover, we designed a novel focal domain generalization loss (FDG-loss) function to enhance the target sample weight and to learn domain consistency features. Furthermore, to enhance the interpretability and flexibility of SEEG-Net, we explain SEEG-Net from multiple perspectives, such as significantly different features, interpretable models, and model learning process interpretation by Grad-CAM++. RESULTS: The performance of our proposed method is verified on a public benchmark multicenter SEEG dataset and a private clinical SEEG dataset for a robust comparison. The experimental results demonstrate that the SEEG-Net model achieves the highest sensitivity and is state-of-the-art on cross-subject (for different patients) evaluation, and well deal with the known problems. Besides, we provide an SEEG processing and database construction flow, by maintaining consistency with the real-world clinical scenarios. SIGNIFICANCE: According to the results, SEEG-Net is constructed to increase the sensitivity of SEEG pathological activity detection. Simultaneously, we settled certain problems about AI assistance in clinical DRE, built a bridge between AI algorithm application and clinical practice.
Asunto(s)
Aprendizaje Profundo , Epilepsia Refractaria , Inteligencia Artificial , Electroencefalografía , Humanos , Técnicas EstereotáxicasRESUMEN
Accurate identification of high-frequency oscillation (HFO) is an important prerequisite for precise localization of epileptic foci and good prognosis of drug-refractory epilepsy. Exploring a high-performance automatic detection method for HFOs can effectively help clinicians reduce the error rate and reduce manpower. Due to the limited analysis perspective and simple model design, it is difficult to meet the requirements of clinical application by the existing methods. Therefore, an end-to-end bi-branch fusion model is proposed to automatically detect HFOs. With the filtered band-pass signal (signal branch) and time-frequency image (TFpic branch) as the input of the model, two backbone networks for deep feature extraction are established, respectively. Specifically, a hybrid model based on ResNet1d and long short-term memory (LSTM) is designed for signal branch, which can focus on both the features in time and space dimension, while a ResNet2d with a Convolutional Block Attention Module (CBAM) is constructed for TFpic branch, by which more attention is paid to useful information of TF images. Then the outputs of two branches are fused to realize end-to-end automatic identification of HFOs. Our method is verified on 5 patients with intractable epilepsy. In intravalidation, the proposed method obtained high sensitivity of 94.62%, specificity of 92.7%, and F1-score of 93.33%, and in cross-validation, our method achieved high sensitivity of 92.00%, specificity of 88.26%, and F1-score of 89.11% on average. The results show that the proposed method outperforms the existing detection paradigms of either single signal or single time-frequency diagram strategy. In addition, the average kappa coefficient of visual analysis and automatic detection results is 0.795. The method shows strong generalization ability and high degree of consistency with the gold standard meanwhile. Therefore, it has great potential to be a clinical assistant tool.
Asunto(s)
Electroencefalografía , Epilepsia , Recolección de Datos , Epilepsia/diagnóstico , Humanos , Redes Neurales de la Computación , Proyectos de InvestigaciónRESUMEN
Surgical intervention or the control of drug-refractory epilepsy requires accurate analysis of invasive inspection intracranial EEG (iEEG) data. A multi-branch deep learning fusion model is proposed to identify epileptogenic signals from the epileptogenic area of the brain. The classical approach extracts multi-domain signal wave features to construct a time-series feature sequence and then abstracts it through the bi-directional long short-term memory attention machine (Bi-LSTM-AM) classifier. The deep learning approach uses raw time-series signals to build a one-dimensional convolutional neural network (1D-CNN) to achieve end-to-end deep feature extraction and signal detection. These two branches are integrated to obtain deep fusion features and results. Resampling is employed to split the imbalanced epileptogenic and non-epileptogenic samples into balanced subsets for clinical validation. The model is validated over two publicly available benchmark iEEG databases to verify its effectiveness on a private, large-scale, clinical stereo EEG database. The model achieves high sensitivity (97.78%), accuracy (97.60%), and specificity (97.42%) on the Bern-Barcelona database, surpassing the performance of existing state-of-the-art techniques. It is then demonstrated on a clinical dataset with an average intra-subject accuracy of 92.53% and cross-subject accuracy of 88.03%. The results suggest that the proposed method is a valuable and extremely robust approach to help researchers and clinicians develop an automated method to identify the source of iEEG signals.