Unsupervised Learning for Lesion Detection on 7T Brain MRI in Patients with Focal Epilepsy
Poster No:
1402
Submission Type:
Abstract Submission
Authors:
Soumen Ghosh1,2, Viktor Vegh1,2, John Phamnguyen1,3, Shahrzad Moinian1,2, David Reutens1,3
Institutions:
1Centre for Advanced Imaging, The University of Queensland, Brisbane, Australia, 2ARC Centre for Innovation in Biomedical Imaging Technology, The University of Queensland, Brisbane, Australia, 3Royal Brisbane and Women’s Hospital, Brisbane, Australia
First Author:
Soumen Ghosh
Centre for Advanced Imaging, The University of Queensland|ARC Centre for Innovation in Biomedical Imaging Technology, The University of Queensland
Brisbane, Australia|Brisbane, Australia
Centre for Advanced Imaging, The University of Queensland|ARC Centre for Innovation in Biomedical Imaging Technology, The University of Queensland
Brisbane, Australia|Brisbane, Australia
Co-Author(s):
Viktor Vegh, Associate Professor
Centre for Advanced Imaging, The University of Queensland|ARC Centre for Innovation in Biomedical Imaging Technology, The University of Queensland
Brisbane, Australia|Brisbane, Australia
Centre for Advanced Imaging, The University of Queensland|ARC Centre for Innovation in Biomedical Imaging Technology, The University of Queensland
Brisbane, Australia|Brisbane, Australia
John Phamnguyen, PhD Candidate
Centre for Advanced Imaging, The University of Queensland|Royal Brisbane and Women’s Hospital
Brisbane, Australia|Brisbane, Australia
Centre for Advanced Imaging, The University of Queensland|Royal Brisbane and Women’s Hospital
Brisbane, Australia|Brisbane, Australia
Shahrzad Moinian, Postdoc
Centre for Advanced Imaging, The University of Queensland|ARC Centre for Innovation in Biomedical Imaging Technology, The University of Queensland
Brisbane, Australia|Brisbane, Australia
Centre for Advanced Imaging, The University of Queensland|ARC Centre for Innovation in Biomedical Imaging Technology, The University of Queensland
Brisbane, Australia|Brisbane, Australia
David Reutens, Emeritus Professor
Centre for Advanced Imaging, The University of Queensland|Royal Brisbane and Women’s Hospital
Brisbane, Australia|Brisbane, Australia
Centre for Advanced Imaging, The University of Queensland|Royal Brisbane and Women’s Hospital
Brisbane, Australia|Brisbane, Australia
Introduction:
The localization of a focal epileptogenic lesion on MRI by visual inspection is often challenging and even when a lesion is identified its delineation using current approaches can be incomplete. These considerations motivated the development of automated methods for lesion identification based on machine learning approaches.
Methods:
The study was approved by the Human Research Ethics Committee Queensland (HREC/17/QRBW/284), and focal epilepsy patients were recruited from the Epilepsy Clinic at the Royal Brisbane Women's Hospital (Brisbane, Australia). MRI scans were performed in-house at the Centre for Advanced Imaging, The University of Queensland. 3D GRE-MRI flow compensated scans were obtained using a 7T ultra-high field whole-body MRI research scanner (Siemens Healthcare, Erlangen, Germany) equipped with a single channel transmit and 32 channel receive head coil (Nova Medical, Wilmington, USA) using voxel size = 0.75 × 0.75 × 0.75 mm3. MP2RAGE data with the same resolution were also acquired.
7T MRI scans of the brain were acquired in patients with focal epilepsy and in healthy controls. Pre-processing of the T1 and T2-weighted images comprised skull stripping using SPM (Version 7.4) followed by the creation of an FSL brain mask (Version 6.0.7) and registration to the MNI 1 mm3 brain template using MIPAV (Version 10.0.0). A one-class support vector machine (oc-SVM) was the constructed.
Figure 1 shows two basic building blocks: unsupervised learning-based feature representation and anomaly detection. The first block extracts a latent representation of the patch from brain MRI using a Siamese autoencoder consisting of convolution and deconvolution layers. We implemented the state-of-the-art (SOA) method,1 and two machine learning methods, all of which use images patches as the input:
• SOA: 2D patch-based autoencoder with input size 15×15×2 comprising two 15×15 obtained from T1 and T2-weighted MRI.
• 2D patch-based autoencoder: Same as SOA with static (i.e., epoch independent) regularization incorporated into the loss function.
• 2.5D patch-based autoencoder: SOA with static regularization and increased the input patch dimension to 15×15×6. Here, T1 and T2-weighted MRI patches are added from the other two orthogonal dimensions to increase the number of patches at the input from 2 to 6. In this case the autoencoder is able to extract information based on 3D brain information.
The anomaly detection module employs an oc-SVM trained using the 64-entry feature vector extracted from either the SOA, 2D and 2.5D Siamese autoencoders.
An oc-SVM model was trained and tested for each image voxel coordinate. Training utilised data from 62 healthy subjects by taking patches from two different healthy subjects at the same time. Data from 5 epilepsy patients were used for testing. The output was a score map of abnormal voxels. Model performance was evaluated by comparing the score map with the ground truth MRI, delineated on MRI by an expert neurologist aided by clinical findings, SPECT, and PET images. Lesion detection accuracy was calculated using Positive Predicted Value (PPV), Dice Index Score (DSC) and Jaccard Similarity Index between the predicted lesion and Ground Truth.
7T MRI scans of the brain were acquired in patients with focal epilepsy and in healthy controls. Pre-processing of the T1 and T2-weighted images comprised skull stripping using SPM (Version 7.4) followed by the creation of an FSL brain mask (Version 6.0.7) and registration to the MNI 1 mm3 brain template using MIPAV (Version 10.0.0). A one-class support vector machine (oc-SVM) was the constructed.
Figure 1 shows two basic building blocks: unsupervised learning-based feature representation and anomaly detection. The first block extracts a latent representation of the patch from brain MRI using a Siamese autoencoder consisting of convolution and deconvolution layers. We implemented the state-of-the-art (SOA) method,1 and two machine learning methods, all of which use images patches as the input:
• SOA: 2D patch-based autoencoder with input size 15×15×2 comprising two 15×15 obtained from T1 and T2-weighted MRI.
• 2D patch-based autoencoder: Same as SOA with static (i.e., epoch independent) regularization incorporated into the loss function.
• 2.5D patch-based autoencoder: SOA with static regularization and increased the input patch dimension to 15×15×6. Here, T1 and T2-weighted MRI patches are added from the other two orthogonal dimensions to increase the number of patches at the input from 2 to 6. In this case the autoencoder is able to extract information based on 3D brain information.
The anomaly detection module employs an oc-SVM trained using the 64-entry feature vector extracted from either the SOA, 2D and 2.5D Siamese autoencoders.
An oc-SVM model was trained and tested for each image voxel coordinate. Training utilised data from 62 healthy subjects by taking patches from two different healthy subjects at the same time. Data from 5 epilepsy patients were used for testing. The output was a score map of abnormal voxels. Model performance was evaluated by comparing the score map with the ground truth MRI, delineated on MRI by an expert neurologist aided by clinical findings, SPECT, and PET images. Lesion detection accuracy was calculated using Positive Predicted Value (PPV), Dice Index Score (DSC) and Jaccard Similarity Index between the predicted lesion and Ground Truth.
·Proposed machine learning framework
Results:
All models were tested on five patients, and all models detected a lesion in four patients. For the four positive lesion patients, Figure 2 depicts the 7T brain MRI, the ground truth, and lesions identified by the SOA and 2D implementations. Lesion detection accuracy for the four patients is summarized in Table 1. Generally, the 2.5D model outperformed the 2D implementation and the SOA method.
·Results
Conclusions:
The proposed 2.5D lesion detection model may provide an automated way to identify MRI-positive focal epilepsy lesions using T1 and T2 weighted MRI.
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Methods Development 2
Keywords:
Computational Neuroscience
Design and Analysis
DISORDERS
Epilepsy
Machine Learning
MRI
Neurological
Source Localization
STRUCTURAL MRI
1|2Indicates the priority used for review