Assessing the Accuracy and Reproducibility of PARIETAL: A Deep Learning Brain Extraction Algorithm.

BACKGROUND: Manual brain extraction from magnetic resonance (MR) images is time-consuming and prone to intra- and inter-rater variability. Several automated approaches have been developed to alleviate these constraints, including deep learning pipelines. However, these methods tend to reduce their performance in unseen magnetic resonance imaging (MRI) scanner vendors and different imaging protocols. PURPOSE: To present and evaluate for clinical use PARIETAL, a pre-trained deep learning brain extraction method. We compare its reproducibility in a scan/rescan analysis and its robustness among scanners of different manufacturers. STUDY TYPE: Retrospective. POPULATION: Twenty-one subjects (12 women) with age range 22-48 years acquired using three different MRI scanner machines including scan/rescan in each of them. FIELD STRENGTH/SEQUENCE: T1-weighted images acquired in a 3-T Siemens with magnetization prepared rapid gradient-echo sequence and two 1.5 T scanners, Philips and GE, with spin-echo and spoiled gradient-recalled (SPGR) sequences, respectively. ASSESSMENT: Analysis of the intracranial cavity volumes obtained for each subject on the three different scanners and the scan/rescan acquisitions. STATISTICAL TESTS: Parametric permutation tests of the differences in volumes to rank and statistically evaluate the performance of PARIETAL compared to state-of-the-art methods. RESULTS: The mean absolute intracranial volume differences obtained by PARIETAL in the scan/rescan analysis were 1.88 mL, 3.91 mL, and 4.71 mL for Siemens, GE, and Philips scanners, respectively. PARIETAL was the best-ranked method on Siemens and GE scanners, while decreasing to Rank 2 on the Philips images. Intracranial differences for the same subject between scanners were 5.46 mL, 27.16 mL, and 30.44 mL for GE/Philips, Siemens/Philips, and Siemens/GE comparison, respectively. The permutation tests revealed that PARIETAL was always in Rank 1, obtaining the most similar volumetric results between scanners. DATA CONCLUSION: PARIETAL accurately segments the brain and it generalizes to images acquired at different sites without the need of training or fine-tuning it again. PARIETAL is publicly available. LEVEL OF EVIDENCE: 2 TECHNICAL EFFICACY STAGE: 2.

Minimizing the effect of white matter lesions on deep learning based tissue segmentation for brain volumetry.

Automated methods for segmentation-based brain volumetry may be confounded by the presence of white matter (WM) lesions, which introduce abnormal intensities that can alter the classification of not only neighboring but also distant brain tissue. These lesions are common in pathologies where brain volumetry is also an important prognostic marker, such as in multiple sclerosis (MS), and thus reducing their effects is critical for improving volumetric accuracy and reliability. In this work, we analyze the effect of WM lesions on deep learning based brain tissue segmentation methods for brain volumetry and introduce techniques to reduce the error these lesions produce on the measured volumes. We propose a 3D patch-based deep learning framework for brain tissue segmentation which is trained on the outputs of a reference classical method. To deal more robustly with pathological cases having WM lesions, we use a combination of small patches and a percentile-based input normalization. To minimize the effect of WM lesions, we also propose a multi-task double U-Net architecture performing end-to-end inpainting and segmentation, along with a training data generation procedure. In the evaluation, we first analyze the error introduced by artificial WM lesions on our framework as well as in the reference segmentation method without the use of lesion inpainting techniques. To the best of our knowledge, this is the first analysis of WM lesion effect on a deep learning based tissue segmentation approach for brain volumetry. The proposed framework shows a significantly smaller and more localized error introduced by WM lesions than the reference segmentation method, that displays much larger global differences. We also evaluated the proposed lesion effect minimization technique by comparing the measured volumes before and after introducing artificial WM lesions to healthy images. The proposed approach performing end-to-end inpainting and segmentation effectively reduces the error introduced by small and large WM lesions in the resulting volumetry, obtaining absolute volume differences of 0.01 ± 0.03% for GM and 0.02 ± 0.04% for WM. Increasing the accuracy and reliability of automated brain volumetry methods will reduce the sample size needed to establish meaningful correlations in clinical studies and allow its use in individualized assessments as a diagnostic and prognostic marker for neurodegenerative pathologies.

Evaluating the use of synthetic T1-w images in new T2 lesion detection in multiple sclerosis.

The assessment of disease activity using serial brain MRI scans is one of the most valuable strategies for monitoring treatment response in patients with multiple sclerosis (MS) receiving disease-modifying treatments. Recently, several deep learning approaches have been proposed to improve this analysis, obtaining a good trade-off between sensitivity and specificity, especially when using T1-w and T2-FLAIR images as inputs. However, the need to acquire two different types of images is time-consuming, costly and not always available in clinical practice. In this paper, we investigate an approach to generate synthetic T1-w images from T2-FLAIR images and subsequently analyse the impact of using original and synthetic T1-w images on the performance of a state-of-the-art approach for longitudinal MS lesion detection. We evaluate our approach on a dataset containing 136 images from MS patients, and 73 images with lesion activity (the appearance of new T2 lesions in follow-up scans). To evaluate the synthesis of the images, we analyse the structural similarity index metric and the median absolute error and obtain consistent results. To study the impact of synthetic T1-w images, we evaluate the performance of the new lesion detection approach when using (1) both T2-FLAIR and T1-w original images, (2) only T2-FLAIR images, and (3) both T2-FLAIR and synthetic T1-w images. Sensitivities of 0.75, 0.63, and 0.81, respectively, were obtained at the same false-positive rate (0.14) for all experiments. In addition, we also present the results obtained when using the data from the international MSSEG-2 challenge, showing also an improvement when including synthetic T1-w images. In conclusion, we show that the use of synthetic images can support the lack of data or even be used instead of the original image to homogenize the contrast of the different acquisitions in new T2 lesions detection algorithms.

Hemorrhagic stroke lesion segmentation using a 3D U-Net with squeeze-and-excitation blocks.

Hemorrhagic stroke is the condition involving the rupture of a vessel inside the brain and is characterized by high mortality rates. Even if the patient survives, stroke can cause temporary or permanent disability depending on how long blood flow has been interrupted. Therefore, it is crucial to act fast to prevent irreversible damage. In this work, a deep learning-based approach to automatically segment hemorrhagic stroke lesions in CT scans is proposed. Our approach is based on a 3D U-Net architecture which incorporates the recently proposed squeeze-and-excitation blocks. Moreover, a restrictive patch sampling is proposed to alleviate the class imbalance problem and also to deal with the issue of intra-ventricular hemorrhage, which has not been considered as a stroke lesion in our study. Moreover, we also analyzed the effect of patch size, the use of different modalities, data augmentation and the incorporation of different loss functions on the segmentation results. All analyses have been performed using a five fold cross-validation strategy on a clinical dataset composed of 76 cases. Obtained results demonstrate that the introduction of squeeze-and-excitation blocks, together with the restrictive patch sampling and symmetric modality augmentation, significantly improved the obtained results, achieving a mean DSC of 0.86±0.074, showing promising automated segmentation results.

Acute and sub-acute stroke lesion segmentation from multimodal MRI.

BACKGROUND AND OBJECTIVE: Acute stroke lesion segmentation tasks are of great clinical interest as they can help doctors make better informed time-critical treatment decisions. Magnetic resonance imaging (MRI) is time demanding but can provide images that are considered the gold standard for diagnosis. Automated stroke lesion segmentation can provide with an estimate of the location and volume of the lesioned tissue, which can help in the clinical practice to better assess and evaluate the risks of each treatment. METHODS: We propose a deep learning methodology for acute and sub-acute stroke lesion segmentation using multimodal MR imaging. We pre-process the data to facilitate learning features based on the symmetry of brain hemispheres. The issue of class imbalance is tackled using small patches with a balanced training patch sampling strategy and a dynamically weighted loss function. Moreover, a combination of whole patch predictions, using a U-Net based CNN architecture, and high degree of overlapping patches reduces the need for additional post-processing. RESULTS: The proposed method is evaluated using two public datasets from the 2015 Ischemic Stroke Lesion Segmentation challenge (ISLES 2015). These involve the tasks of sub-acute stroke lesion segmentation (SISS) and acute stroke penumbra estimation (SPES) from multiple diffusion, perfusion and anatomical MRI modalities. The performance is compared against state-of-the-art methods with a blind online testing set evaluation on each of the challenges. At the time of submitting this manuscript, our approach is the first method in the online rankings for the SISS (DSC=0.59 ± 0.31) and SPES sub-tasks (DSC=0.84 ± 0.10). When compared with the rest of submitted strategies, we achieve top rank performance with a lower Hausdorff distance. CONCLUSIONS: Better segmentation results are obtained by leveraging the anatomy and pathophysiology of acute stroke lesions and using a combined approach to minimize the effects of class imbalance. The same training procedure is used for both tasks, showing the proposed methodology can generalize well enough to deal with different unrelated tasks and imaging modalities without hyper-parameter tuning. In order to promote the reproducibility of our results, a public version of the proposed method has been released to the scientific community.

Acute ischemic stroke lesion core segmentation in CT perfusion images using fully convolutional neural networks.

The use of Computed Tomography (CT) imaging for patients with stroke symptoms is an essential step for triaging and diagnosis in many hospitals. However, the subtle expression of ischemia in acute CT images has made it hard for automated methods to extract potentially quantifiable information. In this work, we present and evaluate an automated deep learning tool for acute stroke lesion core segmentation from CT and CT perfusion images. For evaluation, the Ischemic Stroke Lesion Segmentation (ISLES) 2018 challenge dataset is used that includes 94 cases for training and 62 for testing. The presented method is an improved version of our workshop challenge approach that was ranked among the workshop challenge finalists. The introduced contributions include a more regularized network training procedure, symmetric modality augmentation and uncertainty filtering. Each of these steps is quantitatively evaluated by cross-validation on the training set. Moreover, our proposal is evaluated against other state-of-the-art methods with a blind testing set evaluation using the challenge website, which maintains an ongoing leaderboard for fair and direct method comparison. The tool reaches competitive performance ranking among the top performing methods of the ISLES 2018 testing leaderboard with an average Dice similarity coefficient of 49%. In the clinical setting, this method can provide an estimate of lesion core size and location without performing time costly magnetic resonance imaging. The presented tool is made publicly available for the research community.