Detection and Segmentation of Enlarged Perivascular Spaces in the Brain
Poster No:
1929
Submission Type:
Abstract Submission
Authors:
Mariam Zabihi1, Carole Sudre1
Institutions:
1University College London (UCL), LONDON, Select a State
First Author:
Co-Author:
Introduction:
Enlarged perivascular spaces (EPVS) are small, fluid-filled cavities in the brain, associated with cognitive decline and dementia. Detecting EPVS is important for understanding cerebral small vessel diseases, but it's challenging due to their small size and variable locations, often confused with white matter hyperintensities in imaging [1-3]. Advances in 3D object detection models in medical imaging have improved the detection of small features like EPVS[4-6]. However, these methods face limitations in accurately capturing the complex morphology of EPVS. This paper presents a model that integrates advanced detection techniques with more representative geometric shapes, aiming to enhance the detection of EPVS. Our approach potentially improves the accuracy in identifying EPVS, contributing to the broader efforts in neurological research and the understanding of cerebrovascular diseases.
Methods:
In our study, we employed a 3D Region Convolutional Neural Network (RCNN) for detecting and segmenting Enlarged Perivascular Spaces (EPVS), involving four stages:
1. Feature Learning: A 3D convolutional U-Net architecture [7]was trained on input images to learn key features, using a distance map to enhance feature extraction.
2. Region Proposal Network (RPN): Feature maps from the U-Net fed into an RPN, which generated score maps to identify potential EPVS-containing regions.
3. Non-Maximum Suppression: This stage involved filtering low-confidence regions and eliminating overlapping ellipsoids, using Hellinger [8] and Euclidean distances for optimization.
4. Object Detection and Segmentation: High-confidence ellipsoids were analyzed by another network for precise object detection, integrated with the U-Net's distance map for improved segmentation.
For object shape encoding, a simplified seven-parameter ellipsoid was used, focusing on the largest eigenvalue for scale and direction, and fractional anisotropy for spread.
The method's pipeline is shown in Figure 1.
For performance assessment in our study, we used the VALDO challenge code, providing four key metrics: two for detection (Absolute element difference and F1 score) and two for segmentation quality (Mean dice over true positive elements and absolute volume difference). An element was deemed a true positive if its Intersection over Union with the ground truth exceeded 0.10. Our experiments focused on:
- Segmentation performance based on different imaging modalities (T1, T2, FLAIR, and their combinations).
- Performance variability in segmentation and detection based on brain location.
- Performance on the VALDO test set (SABRE component).
We trained seven distinct models using different modality combinations and assessed them using leave-one-out cross-validation. Each model was trained on random 3D image patches.
Data: The Where is VALDO? challenge was run in 2021 as satellite event of MICCAI [9]. It featured 3 tasks focusing on the detection and segmentation of small markers of cerebral small vessel disease namely EPVS, cerebral microbleeds, and lacunes. For this study, we used the six subjects from the SABRE dataset available for training. The elements were segmented manually using structural MR sequences coregistered to the 1mm3 isotropic T1-weighted sequence (T1-weighted, T2-weighted, T2 FLAIR) by two raters. The final label was generated by taking the union of the objects that have been annotated by both raters. Furthermore, only objects larger than 2 voxels were considered, resulting in a database comprising 1864 EPVS elements.
1. Feature Learning: A 3D convolutional U-Net architecture [7]was trained on input images to learn key features, using a distance map to enhance feature extraction.
2. Region Proposal Network (RPN): Feature maps from the U-Net fed into an RPN, which generated score maps to identify potential EPVS-containing regions.
3. Non-Maximum Suppression: This stage involved filtering low-confidence regions and eliminating overlapping ellipsoids, using Hellinger [8] and Euclidean distances for optimization.
4. Object Detection and Segmentation: High-confidence ellipsoids were analyzed by another network for precise object detection, integrated with the U-Net's distance map for improved segmentation.
For object shape encoding, a simplified seven-parameter ellipsoid was used, focusing on the largest eigenvalue for scale and direction, and fractional anisotropy for spread.
The method's pipeline is shown in Figure 1.
For performance assessment in our study, we used the VALDO challenge code, providing four key metrics: two for detection (Absolute element difference and F1 score) and two for segmentation quality (Mean dice over true positive elements and absolute volume difference). An element was deemed a true positive if its Intersection over Union with the ground truth exceeded 0.10. Our experiments focused on:
- Segmentation performance based on different imaging modalities (T1, T2, FLAIR, and their combinations).
- Performance variability in segmentation and detection based on brain location.
- Performance on the VALDO test set (SABRE component).
We trained seven distinct models using different modality combinations and assessed them using leave-one-out cross-validation. Each model was trained on random 3D image patches.
Data: The Where is VALDO? challenge was run in 2021 as satellite event of MICCAI [9]. It featured 3 tasks focusing on the detection and segmentation of small markers of cerebral small vessel disease namely EPVS, cerebral microbleeds, and lacunes. For this study, we used the six subjects from the SABRE dataset available for training. The elements were segmented manually using structural MR sequences coregistered to the 1mm3 isotropic T1-weighted sequence (T1-weighted, T2-weighted, T2 FLAIR) by two raters. The final label was generated by taking the union of the objects that have been annotated by both raters. Furthermore, only objects larger than 2 voxels were considered, resulting in a database comprising 1864 EPVS elements.
·Figure 1: Methods' overview
Results:
The out-of-sample results presented in Table 1 show the average performance across the cross-validated models for the different modality combinations. Figure 2 shows an example of the gold standard and the model's output in T2 modality.
·Figure 2 and Table 1
Conclusions:
his study adapts advanced object detection methods, using ellipsoid shapes in Fast R-CNN, to improve EPVS detection. Results show that T2 imaging, alone or with T1, enhances detection and segmentaion performance.
Modeling and Analysis Methods:
Methods Development 1
Segmentation and Parcellation
Novel Imaging Acquisition Methods:
Anatomical MRI 2
Keywords:
Other - RCNN;Deep learning;object detection;EPVS;MRI
1|2Indicates the priority used for review
Provide references using author date format
the Aging Brain’. Neurology 98(3) 107-117
2. Paradise M. (2021), ‘ Association of Dilated Perivascular Spaces With Cognitive Decline and
Incident Dementia’, Neurology 96(11) 1501-1511
3. Ding J. (2017),’ Large Perivascular Spaces Visible on Magnetic Resonance Imaging, Cerebral Small Vessel Disease Progression, and Risk of Dementia: The Age, Gene/Environment Susceptibility–Reykjavik Study. JAMA Neurology74(9)1105–1112 (2017)
4. sgari Taghanaki S.(2021),’ Deep semantic segmentation of natural and medical images: a review. Artificial Intelligence’, Review54 137–178
5. Ranjbarzadeh R. (2023),’Brain tumor segmentation of MRI images: A comprehensive re-
view on the application of artificial intelligence tools’. Computers in Biology and Medicine 152
6. Ribli D. (2018),’ Detecting and classifying lesions in mammograms with Deep Learning’.
Scientific Reports8(1: 4165
7. Çiçek, Ö.(2016),’ 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Anno-
tation’ . Medical Image Computing and Computer-Assisted Intervention – MICCAI
8. Fu, GH.:(2020). Hellinger distance-based stable sparse feature selection for high-
dimensional class-imbalanced data. BMC Bioinformatics 21(121)
9. Sudre, C. (2022). Where is VALDO? VAscular lesions detection and segmentation
challenge at MICCAI 2021. arXiv preprint arXiv:2208.07167