Comparison and validation of seven white matter hyperintensities segmentation software in elderly patients.

BACKGROUND: Manual segmentation is currently the gold standard to assess white matter hyperintensities (WMH), but it is time consuming and subject to intra and inter-operator variability. PURPOSE: To compare automatic methods to segment white matter hyperintensities (WMH) in the elderly in order to assist radiologist and researchers in selecting the most relevant method for application on clinical or research data. MATERIAL AND METHODS: We studied a research dataset composed of 147 patients, including 97 patients from the Alzheimer's Disease Neuroimaging Initiative (ADNI) 2 database and 50 patients from ADNI 3 and a clinical routine dataset comprising 60 patients referred for cognitive impairment at the Pitié-Salpêtrière hospital (imaged using four different MRI machines). We used manual segmentation as the gold standard reference. Both manual and automatic segmentations were performed using FLAIR MRI. We compared seven freely available methods that produce segmentation mask and are usable by a radiologist without a strong knowledge of computer programming: LGA (Schmidt et al., 2012), LPA (Schmidt, 2017), BIANCA (Griffanti et al., 2016), UBO detector (Jiang et al., 2018), W2MHS (Ithapu et al., 2014), nicMSlesion (with and without retraining) (Valverde et al., 2019, 2017). The primary outcome for assessing segmentation accuracy was the Dice similarity coefficient (DSC) between the manual and the automatic segmentation software. Secondary outcomes included five other metrics. RESULTS: A deep learning approach, NicMSlesion, retrained on data from the research dataset ADNI, performed best on this research dataset (DSC: 0.595) and its DSC was significantly higher than that of all others. However, it ranked fifth on the clinical routine dataset and its performance severely dropped on data with artifacts. On the clinical routine dataset, the three top-ranked methods were LPA, SLS and BIANCA. Their performance did not differ significantly but was significantly higher than that of other methods. CONCLUSION: This work provides an objective comparison of methods for WMH segmentation. Results can be used by radiologists to select a tool.

Convolutional neural networks for classification of Alzheimer's disease: Overview and reproducible evaluation.

Numerous machine learning (ML) approaches have been proposed for automatic classification of Alzheimer's disease (AD) from brain imaging data. In particular, over 30 papers have proposed to use convolutional neural networks (CNN) for AD classification from anatomical MRI. However, the classification performance is difficult to compare across studies due to variations in components such as participant selection, image preprocessing or validation procedure. Moreover, these studies are hardly reproducible because their frameworks are not publicly accessible and because implementation details are lacking. Lastly, some of these papers may report a biased performance due to inadequate or unclear validation or model selection procedures. In the present work, we aim to address these limitations through three main contributions. First, we performed a systematic literature review. We identified four main types of approaches: i) 2D slice-level, ii) 3D patch-level, iii) ROI-based and iv) 3D subject-level CNN. Moreover, we found that more than half of the surveyed papers may have suffered from data leakage and thus reported biased performance. Our second contribution is the extension of our open-source framework for classification of AD using CNN and T1-weighted MRI. The framework comprises previously developed tools to automatically convert ADNI, AIBL and OASIS data into the BIDS standard, and a modular set of image preprocessing procedures, classification architectures and evaluation procedures dedicated to deep learning. Finally, we used this framework to rigorously compare different CNN architectures. The data was split into training/validation/test sets at the very beginning and only the training/validation sets were used for model selection. To avoid any overfitting, the test sets were left untouched until the end of the peer-review process. Overall, the different 3D approaches (3D-subject, 3D-ROI, 3D-patch) achieved similar performances while that of the 2D slice approach was lower. Of note, the different CNN approaches did not perform better than a SVM with voxel-based features. The different approaches generalized well to similar populations but not to datasets with different inclusion criteria or demographical characteristics. All the code of the framework and the experiments is publicly available: general-purpose tools have been integrated into the Clinica software (www.clinica.run) and the paper-specific code is available at: https://github.com/aramis-lab/AD-DL.