Improved Ex-vivo Cerebral Microbleed Detection Using Simulation-Based Self-Supervised Learning
Poster No:
255
Submission Type:
Abstract Submission
Authors:
Grant Nikseresht1, Arnold Evia2, David Bennett2, Julie Schneider2, Gady Agam1, Konstantinos Arfanakis1
Institutions:
1Illinois Institute of Technology, Chicago, IL, 2Rush University Medical Center, Chicago, IL
First Author:
Co-Author(s):
Introduction:
Cerebral microbleed (CMB) annotation on postmortem MRI scans of autopsied brains of community-based older adults is necessary for MR-pathology studies of cerebral small vessel disease (SVD) (Nikseresht 2023, Charidimou 2018). However, automation of CMB detection is challenging due to the low incidence of CMBs in community-based older adult brains and high prevalence of CMB mimics on ex-vivo MRI (Fig. 1). While data synthesis can improve model performance by increasing the amount of available training data, biases in the synthesis model can lead to poor generalization performance. Self-supervised learning (SSL) has been shown to be a powerful tool for improving representation learning in data-scarce environments such as medical imaging (Tang 2022). We propose a novel pretext task called fuzzy segmentation (FuzzSeg) that leverages the data synthesis process as a form of self-supervision. Ex-vivo CMB detection models pre-trained with FuzzSeg are shown to outperform models trained from scratch.
·Figure 1. Overview of the fuzzy segmentation and knowledge transfer using self-supervised learning.
Methods:
286 participants from the Rush Memory and Aging Project (Bennett 2012a) and Religious Orders Study (Bennett 2012b), two longitudinal cohort studies of aging, were included in this work. T2*-weighted gradient echo scans of autopsied brains with a voxel resolution of 1x1x1 mm3 were used after N4 bias correction. CMBs in these images were manually annotated by an experienced rater blinded to all clinical and pathological information.
Given a synthetic example, the goal of fuzzy segmentation is to predict the hidden kernel used to generate it (Fig. 1). The term fuzzy segmentation refers to the fact that each kernel is interpreted as the relative scalar intensity drop in T2* at a particular voxel compared to healthy background. Fuzzy segmentation is useful for pre-training because it requires the model to learn to separate hypointense foreground from background and estimate key features of potential CMBs such as hypointensity shape and relative intensity. Two general-purpose self-supervised pretext tasks, rotation prediction and image inpainting, were also evaluated. A modified 3D ResNet20 backbone was used for feature encoding.Task-specific decoder heads were attached for pre-training. Encoder weights are learned by pre-training on the self-supervised pretext tasks and then transferred to the CMB detection task by replacing the decoder component with a classification head.
An end-to-end CMB detection framework that combines data synthesis, candidate selection, false positive reduction, and full scan evaluation was used as the backbone for this work (Nikseresht 2012b). Input patches of size 16x16x16x4 were used with four signal echoes in the channel dimension. A high-sensitivity candidate selection algorithm was used to identify CMB candidates based on pre-generated image features. Training and evaluation were done using a repeated randomized 5-fold cross-validation technique, and final predictions were generated using ensembling.
Given a synthetic example, the goal of fuzzy segmentation is to predict the hidden kernel used to generate it (Fig. 1). The term fuzzy segmentation refers to the fact that each kernel is interpreted as the relative scalar intensity drop in T2* at a particular voxel compared to healthy background. Fuzzy segmentation is useful for pre-training because it requires the model to learn to separate hypointense foreground from background and estimate key features of potential CMBs such as hypointensity shape and relative intensity. Two general-purpose self-supervised pretext tasks, rotation prediction and image inpainting, were also evaluated. A modified 3D ResNet20 backbone was used for feature encoding.Task-specific decoder heads were attached for pre-training. Encoder weights are learned by pre-training on the self-supervised pretext tasks and then transferred to the CMB detection task by replacing the decoder component with a classification head.
An end-to-end CMB detection framework that combines data synthesis, candidate selection, false positive reduction, and full scan evaluation was used as the backbone for this work (Nikseresht 2012b). Input patches of size 16x16x16x4 were used with four signal echoes in the channel dimension. A high-sensitivity candidate selection algorithm was used to identify CMB candidates based on pre-generated image features. Training and evaluation were done using a repeated randomized 5-fold cross-validation technique, and final predictions were generated using ensembling.
Results:
The CMB detection model jointly pre-trained on fuzzy segmentation and rotation prediction tasks (AP=0.3988) achieved the highest sensitivity at both 0.5 false positives per subject (36.4%) and at 16 false positives per subject (81.5%) of all models evaluated (Fig. 2). Pre-training with fuzzy segmentation alone (AP=0.3748) also led to improvements over a baseline model trained without pre-training (AP=0.3618), pre-training with rotation prediction (AP=0.3721), and pre-training with image inpainting (AP=0.3619).
·Figure 2. Summary of results demonstrating that self-supervised learning with fuzzy segmentation leads to improvements in detection performance.
Conclusions:
This work demonstrates that self-supervised pre-training with FuzzSeg is a data-efficient technique for improving the performance of ex-vivo CMB detection algorithms in community-based cohorts where CMB prevalence is low and mimics are abundant. This has led to reduced labeling time and increased sensitivity for partially automated CMB annotation, a critical step in the development of future MR-pathology studies examining the link between CMBs and neuropathology in community-based older adults.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 1
Lifespan Development:
Aging
Modeling and Analysis Methods:
Methods Development 2
Segmentation and Parcellation
Keywords:
ADULTS
Aging
Cerebrovascular Disease
Machine Learning
MRI
Segmentation
Other - Microbleed
1|2Indicates the priority used for review
Provide references using author date format
Bennett, A. (2012), 'Overview and Findings from the Rush Memory and Aging Project', Current Alzheimer Research, vol. 9, no. 6, pp. 646-663.
Bennett, A. (2012), 'Overview and Findings from the Religious Orders Study', Current Alzheimer Research, vol. 9, no. 6, pp. 628-645.
Charidimou, A. (2018), 'Clinical significance of cerebral microbleeds on MRI: A comprehensive meta-analysis of risk of intracerebral hemorrhage, ischemic stroke, mortality, and dementia in cohort studies (v1)', International Journal of Stroke, vol. 13, no. 5, pp. 454-468.
Nikseresht, G. (2023), 'Neuropathologic correlates of cerebral microbleeds in community-based older adults', Neurobiology of Aging, vol. 129, pp. 89-98.
Nikseresht, G. (2022b), 'Microbleed detection in autopsied brains from community-based older adults using microbleed synthesis and deep learning', Poster presented at: International Society for Magnetic Resonance in Medicine; May 7-12, 2022; London, England.
Tang, Y. (2022), 'Self-Supervised Pre-Training of Swin Transformers for 3D Medical Image Analysis', Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition.