Improved detection of focal cortical dysplasia in normal‐appearing FLAIR images using a Bayesian classifier

• Published in: Medical physics (Lancaster) Vol. 48; no. 2; pp. 912 - 925
• Main Authors: Feng, Cuixia, Zhao, Hulin, Li, Yueer, Cheng, Zhibiao, Wen, Junhai
• Format: Journal Article
• Language: English
• Published: United States 01.02.2021
• Subjects: Bayes Theorem; Bayesian classifier; detection; Epilepsy - diagnostic imaging; FCD; FLAIR‐negative; Humans; Imaging, Three-Dimensional; Magnetic Resonance Imaging; Malformations of Cortical Development - diagnostic imaging; detection; FCD; Bayesian classifier; FLAIR-negative
• ISSN: 0094-2405; 2473-4209; 2473-4209
• DOI: 10.1002/mp.14646

Purpose Focal cortical dysplasia (FCD) is a malformation of cortical development that often causes pharmacologically intractable epilepsy. However, FCD lesions are frequently characterized by minor structural abnormalities that can easily go unrecognized, making diagnosis difficult. Therefore, many epileptic patients have had pathologically confirmed FCD lesions that appeared normal in pre‐surgical fluid‐attenuated inversion recovery (FLAIR) magnetic resonance (MR) studies. Such lesions are called “FLAIR‐negative.” This study aimed to improve the detection of histopathologically verified FCD in a sample of patients without visually appreciable lesions. Methods The technique first extracts a series of features from a FLAIR image. Then, three naive Bayesian classifiers with probability (NBCP) are trained based on different numbers of feature maps to classify voxels as lesional or healthy voxels and assign the lesions a probability of correct classification. This method classifies the three‐dimensional (3D) images of all patients using leave‐one‐out cross‐validation (LOOCV). Finally, the 3D lesion probability map, including epileptogenic lesions, is obtained by removing false‐positive voxel outliers using the morphological method. The performance of the NBCP was assessed for quantitative analysis by specificity, accuracy, recall, precision, and Dice coefficient in subject‐wise, lesion‐wise, and voxel‐wise manners. Results The best detection results were obtained by using four features: cortical thickness, symmetry, K‐means, and modified texture energy. There were eight lesions in seven patients. The subject‐wise sensitivity of the proposed method was 85.71% (6/7). Seven out of eight lesions were detected, so the lesion‐wise sensitivity was 87.50% (7/8). No significant differences in effectiveness were found between automated lesion detection using four features and lesion detection using manual segmentation, as voxels were quantitatively analyzed in terms of specificity (mean ± SD = 99.64 ± 0.13), accuracy (mean ± SD = 99.62 ± 0.14), recall (mean ± SD = 73.27 ± 26.11), precision (mean ± SD = 11.93 ± 8.16), and Dice coefficient (mean ± SD = 22.82 ± 15.57). Conclusion We developed a novel automatic voxel‐based method to improve the detection of FCD FLAIR‐negative lesions. To the best of our knowledge, this study is the first to detect FCD lesions that appear normal in pre‐surgical 3D high‐resolution FLAIR images alone with a limited number of radiomics features. We optimized the algorithm and selected the best prior probability to improve the detection. For non‐temporal lobe epilepsy (non‐TLE) patients, lesions could be accurately located, although there were still false‐positive areas.