| 其他摘要 | Magnetic resonance imaging (MRI) is a non-invasive radiation-free imagingtechnology. But beyond detecting brain lesions or tumors, comparatively little successhas been attained in identifying brain disorders such as Alzheimer’s disease (AD).Many machine learning algorithms to detect AD have been trained using limitedtraining data, meaning they often generalize poorly when applied to scans frompreviously unseen populations. While deep convolutional networks have shownexcellent performance on image-based classification tasks, their application to brainimaging is limited by the restricted sample size. As more and more MRI brain imagingdatabases become publicly available and deep learning excels in image classification,brain imaging big data combined with deep learning becomes a viable path for MRIbrain imaging to move further toward clinical applications.
In the current study, a retrospective MRI dataset pooled from more than 217sites/scanners constituted the largest brain MRI sample to date (85,721 scans from50,876 participants) between January 2017 and August 2021. Next, a state-of-the-artdeep convolutional neural network, Inception-ResNet-V2, was built as a sex classifierwith high generalization capability. The sex classifier achieved 94.9% accuracy andserved as a base model in transfer learning for the objective diagnosis of AD. Theocclusion map showed that hypothalamus, superior vermis and pituitary played criticalroles in predicting sex. To explore the clinical implications of the brain gender classifier,we tested brain gender directly on a homosexual sample, in which 40% of homosexualfemales would be discriminated as brain males. It was a much higher misclassificationrate than the other samples, possibly reflecting some correlation between subjects'sexual orientation and structural brain variation.
We then explored the potential of CNNs for the objective diagnosis of braindiseases. After transfer learning, the model fine-tuned for AD classification achieved91.3% accuracy in leave-sites-out cross-validation on the Alzheimer's DiseaseNeuroimaging Initiative (ADNI, 6,857 samples) dataset and 94.2%/93.6%/90.5%accuracy for direct tests on three unseen independent datasets (The Australian Imaging,Biomarkers & Lifestyle, AIBL, 669 samples / the Minimal Interval Resonance Imagingin Alzheimer's Disease cohort, MIRIAD, 644 samples / Open Access Series of ImagingStudies, OASIS, 1,123 samples). When this AD classifier was tested on brain imagesfrom unseen mild cognitive impairment (MCI) patients, MCI patients who finallyconverted to AD were 3 times more likely to be predicted as AD than MCI patients whodid not convert (65.2% vs 20.6%). Predicted scores from the AD classifier showedsignificant correlations with illness severity. The occlusion map for the AD classifierhighlighted that the hippocampus and parahippocampal gyrus - especially in the lefthemisphere - played unique roles in predicting AD. In sum, the proposed AD classifiercould offer a medical-grade marker that have potential to be integrated into ADdiagnostic practice. In addition, positron emission tomography (PET) with nuclearradiation is commonly used in the diagnosis of AD to assess cerebral metabolism. Theoutput score of our MRI-based AD classifier correlated with the 18F-AV45 PET SUVRvalue of 0.359 (p<0.001), which to some extent suggests the use of non-invasive,radiation-free MRI instead of PET for the assessment of AD. This offers the possibilityof using non-invasive and radiation-free MRI instead of PET scan with radiation in theassessment of AD.
However, the pathology of psychiatric disorders such as MDD, ASD and SCZ isless clear than that of neurological disorder such as AD, the disease causes smallerstructural changes in the brain, and the disease is more heterogeneous, making it moredifficult to establish a disease prediction model based on magnetic resonance images.In addition, the brain disease database's other than AD are generally retrospective indesign, with relatively small data volume, more acquisition sites, and inconsistencies inmodels, sequences, and quality control. These factors lead to greater challenges in thegeneralizability of brain imaging-based classifiers for MDD, ASD, and SCZ. We usedhyperparameters obtained in AD transfer learning to attempt to construct diseaseclassifiers for MDD, ASD, and SCZ using transfer learning, in conjunction with datanormalization methods. For the MDD classifier, we used retrospectively designedshared data from the REST-meta-MDD data consortium of more than ten hospitalscontaining 1,300 MDD patients and 1,128 normal control subjects (NC). For the ASDand SCZ classifiers, we used prospectively designed study data from five hospitals andresearch units that were pre-serialized and harmonized, containing 234 ASD patients,186 SCZ patients, and 128 NC subjects. The results showed that MDD-NC, ASD-NC,SCZ-NC, and ASD-SCZ reached an average accuracy of 55.6%, 39.9%, 55.7%, and48.3% on cross-site validation, with AUCs of 0.562, 0.293, 0.553, and 0.527. We foundthat: 1. Disease classifiers for MDD, ASD, and SCZ performed inferiorly to ADclassifiers, which may reflect the influence of multiple factors such as sample size,disease heterogeneity, and different structural changes caused by disease; 2. Modelsusing random cross-validation performed much better than those using cross-site-validation (MDD-NC, ASD-NC, SCZ-NC, and ASD-SCZ achieved average accuracyof 69.3%, 68.8%, 77.8%, and 74.1% on random cross-validation, with AUCs of 0.770,0.670, 0.842, and 0.805), reflecting the importance of site-effects for classifiers basedon magnetic resonance brain imaging. And a truly clinically useful diagnostic aid mustrequire good generalizability, i.e., stable high-level performance across differentmodels and sequences. In addition, data standardization has a very important impact onimproving the cross-site prediction ability of models, but the common standardizationmethods in the field do not effectively improve the accuracy, although they can makethe brain imaging feature maps more satisfy the machine learning assumption ofindependent homogeneous distribution. These results raise new methodological issuesfor objective assisted diagnosis based on MR brain imaging, and the solution of theseissues in the future is expected to further promote the clinical application of MR brainimaging.
In summary, this study applied and pooled almost all MRI brain imaging publicdatasets to construct a convolutional neural network brain image gender classifier withideal accuracy and generalizability; subsequently, the convolutional neural networkgender classifier was transferred to AD samples using transfer learning to achieve thepractical performance of AD prediction with the feasibility of objective brain imaging-based diagnosis of neurological diseases. The obtained hyperparameters were then used for further transferred to examine the possibility of their application in psychiatricdiseases, but the objective prediction of brain images of psychiatric diseases was notsuccessfully achieved, and further exploration was needed. |
修改评论