{"id":{"repo_id":"cambridge","oai_identifier":"oai:www.repository.cam.ac.uk:1810/395952"},"canonical_url":"https://search.dev.ndltd.org/etd/cambridge/oai:www.repository.cam.ac.uk:1810/395952","repository":{"repo_id":"cambridge","name":"Cambridge University","base_url":"https://api.repository.cam.ac.uk/server/oai/request"},"display":{"title":"Applications of Machine Learning to Neuroimage Analysis in Cerebral Small Vessel Disease","abstract":"Introduction Cerebral small vessel disease (CSVD) causes around 25% of ischaemic strokes and contributes to nearly half of dementia cases worldwide. Neuroimaging, particularly magnetic resonance imaging (MRI), is fundamental to the assessment, prognosis, and pathophysiological study of CSVD. Meanwhile, machine learning (ML), including deep learning (DL), methods are particularly well suited for analysing high-dimensional data such as MRI images. This thesis aimed to investigate how machine learning can be applied to neuroimage analysis in CSVD research, focusing on advancing both CSVD assessment and prognosis. Specifically, I explored how ML can be applied to automate the morphology quantification of lenticulostriate arteries (LSA), which are impacted in CSVD, from 7T MRI (Part II), and how ML can improve the prediction of incident dementia for CSVD patients (Part III). Methods In Part II of this thesis, data from the CamSVD study was used for developing a semi-automated pipeline for segmenting the LSAs and quantifying their 3D morphology from time-of-flight magnetic resonance angiography (TOF-MRA). A two-staged approach was adopted: in the vessel segmentation stage, three leading DL and non-DL segmentation methods were evaluated, namely the DS6 DL model, the nnU-Net DL model, and the classical Frangi filter-based pipeline, MSFDF; in the LSA quantification stage, the branch counts, length, and tortuosity of LSAs within their perfusion area were extracted based on LSA segmentation. This pipeline was validated through comparing the extracted 3D metrics with those from manual analysis measuring LSA morphology on 2D coronal maximum intensity projection (MIP) images, which is currently the standard approach. Finally, this pipeline was applied to the complete CamSVD cohort to investigate the clinical significance of the extracted metrics. In Part III of this thesis, data from three cohorts with varying CSVD severity was initially pooled for investigating whether common ML methods can improve dementia prediction in CSVD compared with traditional statistical approaches, when using established CSVD imaging markers, along with clinical and cognitive predictors. Both survival and classification analyses were performed, and the feature importance for each model was analysed. Furthermore, the imaging data from two of the three cohorts were gathered for investigating whether DL survival models leveraging the original MRI images can improve dementia prediction in CSVD compared with using established CSVD imaging markers. Specifically, T1, FLAIR, and mean diffusivity (MD) images were utilised along with clinical and cognitive features. Both unimodal and multimodal DL survival models were trained with different strategies and compared against Cox regression. Results The major output from Part II was the Lenticulostriate artery Ultra-high-field Morphology Extraction and quantificatioN (LUMEN) pipeline. The finetuned DS6 achieved the best segmentation Dice score (>0.8), and both DS6 and nnU-Net outperformed the classical MSFDF. An easily deployable Jupyter notebook interactively linked to 3D Slicer was developed for LSA quantification. The 3D LSA morphological metrics extracted by LUMEN provide a more faithful characterisation of LSA morphology than those from 2D MIP analysis. However, no significant differences in 3D LSA morphological metrics were found across lacunar stroke patients, non-lacunar stroke controls and healthy controls. The clinical significance of extracted metrics remains to be further investigated. Part III found that when using established CSVD imaging markers along with clinical and cognitive predictors, ML methods performed similarly to Cox and logistic regression, and prediction was dominated by global cognition and age. However, when applied to the original MRI images, DL methods could improve the survival prediction of dementia compared with Cox regression using conventional CSVD imaging markers. MD was the most predictive modality among all, and multimodal DL model trained end to end from scratch achieved the best test c-index. Conclusions In conclusion, this thesis demonstrated that ML can be applied to automate the assessment of 3D LSA morphology for investigating CSVD pathophysiology, and that ML can also improve dementia prediction in CSVD when applied directly to MRI images. Overall, these findings highlight the promising potential of ML-based neuroimage analysis in improving both the assessment and prognosis of CSVD.","abstract_html":"Introduction Cerebral small vessel disease (CSVD) causes around 25% of ischaemic strokes and contributes to nearly half of dementia cases worldwide. Neuroimaging, particularly magnetic resonance imaging (MRI), is fundamental to the assessment, prognosis, and pathophysiological study of CSVD. Meanwhile, machine learning (ML), including deep learning (DL), methods are particularly well suited for analysing high-dimensional data such as MRI images. This thesis aimed to investigate how machine learning can be applied to neuroimage analysis in CSVD research, focusing on advancing both CSVD assessment and prognosis. Specifically, I explored how ML can be applied to automate the morphology quantification of lenticulostriate arteries (LSA), which are impacted in CSVD, from 7T MRI (Part II), and how ML can improve the prediction of incident dementia for CSVD patients (Part III). Methods In Part II of this thesis, data from the CamSVD study was used for developing a semi-automated pipeline for segmenting the LSAs and quantifying their 3D morphology from time-of-flight magnetic resonance angiography (TOF-MRA). A two-staged approach was adopted: in the vessel segmentation stage, three leading DL and non-DL segmentation methods were evaluated, namely the DS6 DL model, the nnU-Net DL model, and the classical Frangi filter-based pipeline, MSFDF; in the LSA quantification stage, the branch counts, length, and tortuosity of LSAs within their perfusion area were extracted based on LSA segmentation. This pipeline was validated through comparing the extracted 3D metrics with those from manual analysis measuring LSA morphology on 2D coronal maximum intensity projection (MIP) images, which is currently the standard approach. Finally, this pipeline was applied to the complete CamSVD cohort to investigate the clinical significance of the extracted metrics. In Part III of this thesis, data from three cohorts with varying CSVD severity was initially pooled for investigating whether common ML methods can improve dementia prediction in CSVD compared with traditional statistical approaches, when using established CSVD imaging markers, along with clinical and cognitive predictors. Both survival and classification analyses were performed, and the feature importance for each model was analysed. Furthermore, the imaging data from two of the three cohorts were gathered for investigating whether DL survival models leveraging the original MRI images can improve dementia prediction in CSVD compared with using established CSVD imaging markers. Specifically, T1, FLAIR, and mean diffusivity (MD) images were utilised along with clinical and cognitive features. Both unimodal and multimodal DL survival models were trained with different strategies and compared against Cox regression. Results The major output from Part II was the Lenticulostriate artery Ultra-high-field Morphology Extraction and quantificatioN (LUMEN) pipeline. The finetuned DS6 achieved the best segmentation Dice score (>0.8), and both DS6 and nnU-Net outperformed the classical MSFDF. An easily deployable Jupyter notebook interactively linked to 3D Slicer was developed for LSA quantification. The 3D LSA morphological metrics extracted by LUMEN provide a more faithful characterisation of LSA morphology than those from 2D MIP analysis. However, no significant differences in 3D LSA morphological metrics were found across lacunar stroke patients, non-lacunar stroke controls and healthy controls. The clinical significance of extracted metrics remains to be further investigated. Part III found that when using established CSVD imaging markers along with clinical and cognitive predictors, ML methods performed similarly to Cox and logistic regression, and prediction was dominated by global cognition and age. However, when applied to the original MRI images, DL methods could improve the survival prediction of dementia compared with Cox regression using conventional CSVD imaging markers. MD was the most predictive modality among all, and multimodal DL model trained end to end from scratch achieved the best test c-index. Conclusions In conclusion, this thesis demonstrated that ML can be applied to automate the assessment of 3D LSA morphology for investigating CSVD pathophysiology, and that ML can also improve dementia prediction in CSVD when applied directly to MRI images. Overall, these findings highlight the promising potential of ML-based neuroimage analysis in improving both the assessment and prognosis of CSVD.","abstract_has_math":false,"creators":["Li, Rui"],"institution":"University of Cambridge","degree_name":"Doctor of Philosophy (PhD)","degree_level":"Doctoral","degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":["Markus, Hugh"],"committee_chairs":[],"committee_members":[],"year":2025,"date_issued":"2025-09-25","date_published":"2025-09-25","updated_at":"2026-07-22T22:23:56Z","subjects":["machine learning","MRI","cerebral small vessel disease"],"languages":["eng"],"rights":[],"rights_urls":["https://www.repository.cam.ac.uk/bitstreams/808c678f-3f26-47c3-bf89-8f1167c09a8b/download","http://purl.org/NET/rdflicense/allrightsreserved"],"identifier_entries":[{"key":"dc:creator.authoridentifier","label":"Author Identifier","values":["0000000305553176"],"render_values":[{"text":"0000-0003-0555-3176","href":"https://orcid.org/0000-0003-0555-3176","code":true}]}]},"links":{"outbound_url":"https://doi.org/10.17863/CAM.125288","outbound_label":"DOI","outbound_source":"dc:identifier.doi"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Markus, Hugh"]},{"key":"dc:creator","label":"Author","values":["Li, Rui"]},{"key":"dc:creator.authoridentifier","label":"Author Identifier","values":["0000000305553176"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.issued","label":"Date","values":["2025-09-25"]},{"key":"dc:publisher.institution","label":"Dc Publisher Institution","values":["University of Cambridge"]},{"key":"dc:relation.isreferencedby.uri","label":"Dc Relation Isreferencedby URI","values":["https://www.repository.cam.ac.uk/handle/1810/395952"]},{"key":"dc:type","label":"Dc Type","values":["Thesis"]},{"key":"dc:type.qualificationlevel","label":"Dc Type Qualificationlevel","values":["Doctoral"]},{"key":"dc:type.qualificationname","label":"Dc Type Qualificationname","values":["Doctor of Philosophy (PhD)"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["machine learning","MRI","cerebral small vessel disease"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["eng"]},{"key":"dc:rights","label":"Dc Rights","values":["https://www.repository.cam.ac.uk/bitstreams/808c678f-3f26-47c3-bf89-8f1167c09a8b/download","http://purl.org/NET/rdflicense/allrightsreserved"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.doi","label":"DOI","values":["https://doi.org/10.17863/CAM.125288"]},{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://www.repository.cam.ac.uk/bitstreams/4feb2db3-9fc6-4c77-908f-b71c1b24e445/download"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["Introduction Cerebral small vessel disease (CSVD) causes around 25% of ischaemic strokes and contributes to nearly half of dementia cases worldwide. Neuroimaging, particularly magnetic resonance imaging (MRI), is fundamental to the assessment, prognosis, and pathophysiological study of CSVD. Meanwhile, machine learning (ML), including deep learning (DL), methods are particularly well suited for analysing high-dimensional data such as MRI images. This thesis aimed to investigate how machine learning can be applied to neuroimage analysis in CSVD research, focusing on advancing both CSVD assessment and prognosis. Specifically, I explored how ML can be applied to automate the morphology quantification of lenticulostriate arteries (LSA), which are impacted in CSVD, from 7T MRI (Part II), and how ML can improve the prediction of incident dementia for CSVD patients (Part III). Methods In Part II of this thesis, data from the CamSVD study was used for developing a semi-automated pipeline for segmenting the LSAs and quantifying their 3D morphology from time-of-flight magnetic resonance angiography (TOF-MRA). A two-staged approach was adopted: in the vessel segmentation stage, three leading DL and non-DL segmentation methods were evaluated, namely the DS6 DL model, the nnU-Net DL model, and the classical Frangi filter-based pipeline, MSFDF; in the LSA quantification stage, the branch counts, length, and tortuosity of LSAs within their perfusion area were extracted based on LSA segmentation. This pipeline was validated through comparing the extracted 3D metrics with those from manual analysis measuring LSA morphology on 2D coronal maximum intensity projection (MIP) images, which is currently the standard approach. Finally, this pipeline was applied to the complete CamSVD cohort to investigate the clinical significance of the extracted metrics. In Part III of this thesis, data from three cohorts with varying CSVD severity was initially pooled for investigating whether common ML methods can improve dementia prediction in CSVD compared with traditional statistical approaches, when using established CSVD imaging markers, along with clinical and cognitive predictors. Both survival and classification analyses were performed, and the feature importance for each model was analysed. Furthermore, the imaging data from two of the three cohorts were gathered for investigating whether DL survival models leveraging the original MRI images can improve dementia prediction in CSVD compared with using established CSVD imaging markers. Specifically, T1, FLAIR, and mean diffusivity (MD) images were utilised along with clinical and cognitive features. Both unimodal and multimodal DL survival models were trained with different strategies and compared against Cox regression. Results The major output from Part II was the Lenticulostriate artery Ultra-high-field Morphology Extraction and quantificatioN (LUMEN) pipeline. The finetuned DS6 achieved the best segmentation Dice score (>0.8), and both DS6 and nnU-Net outperformed the classical MSFDF. An easily deployable Jupyter notebook interactively linked to 3D Slicer was developed for LSA quantification. The 3D LSA morphological metrics extracted by LUMEN provide a more faithful characterisation of LSA morphology than those from 2D MIP analysis. However, no significant differences in 3D LSA morphological metrics were found across lacunar stroke patients, non-lacunar stroke controls and healthy controls. The clinical significance of extracted metrics remains to be further investigated. Part III found that when using established CSVD imaging markers along with clinical and cognitive predictors, ML methods performed similarly to Cox and logistic regression, and prediction was dominated by global cognition and age. However, when applied to the original MRI images, DL methods could improve the survival prediction of dementia compared with Cox regression using conventional CSVD imaging markers. MD was the most predictive modality among all, and multimodal DL model trained end to end from scratch achieved the best test c-index. Conclusions In conclusion, this thesis demonstrated that ML can be applied to automate the assessment of 3D LSA morphology for investigating CSVD pathophysiology, and that ML can also improve dementia prediction in CSVD when applied directly to MRI images. Overall, these findings highlight the promising potential of ML-based neuroimage analysis in improving both the assessment and prognosis of CSVD."]},{"key":"dc:format.checksum.md5","label":"Dc Format Checksum Md5","values":["9f2d2bc51c451e5f4b9725ab074ee4ca","87eda9de84448d1f82354d60eee3eb5f"]},{"key":"dc:title","label":"Title","values":["Applications of Machine Learning to Neuroimage Analysis in Cerebral Small Vessel Disease"]}]}],"canonical_facts":{"dc:contributor.advisor":["Markus, Hugh"],"dc:creator":["Li, Rui"],"dc:creator.authoridentifier":["0000000305553176"],"dc:date.issued":["2025-09-25"],"dc:description.abstract":["Introduction Cerebral small vessel disease (CSVD) causes around 25% of ischaemic strokes and contributes to nearly half of dementia cases worldwide. Neuroimaging, particularly magnetic resonance imaging (MRI), is fundamental to the assessment, prognosis, and pathophysiological study of CSVD. Meanwhile, machine learning (ML), including deep learning (DL), methods are particularly well suited for analysing high-dimensional data such as MRI images. This thesis aimed to investigate how machine learning can be applied to neuroimage analysis in CSVD research, focusing on advancing both CSVD assessment and prognosis. Specifically, I explored how ML can be applied to automate the morphology quantification of lenticulostriate arteries (LSA), which are impacted in CSVD, from 7T MRI (Part II), and how ML can improve the prediction of incident dementia for CSVD patients (Part III). Methods In Part II of this thesis, data from the CamSVD study was used for developing a semi-automated pipeline for segmenting the LSAs and quantifying their 3D morphology from time-of-flight magnetic resonance angiography (TOF-MRA). A two-staged approach was adopted: in the vessel segmentation stage, three leading DL and non-DL segmentation methods were evaluated, namely the DS6 DL model, the nnU-Net DL model, and the classical Frangi filter-based pipeline, MSFDF; in the LSA quantification stage, the branch counts, length, and tortuosity of LSAs within their perfusion area were extracted based on LSA segmentation. This pipeline was validated through comparing the extracted 3D metrics with those from manual analysis measuring LSA morphology on 2D coronal maximum intensity projection (MIP) images, which is currently the standard approach. Finally, this pipeline was applied to the complete CamSVD cohort to investigate the clinical significance of the extracted metrics. In Part III of this thesis, data from three cohorts with varying CSVD severity was initially pooled for investigating whether common ML methods can improve dementia prediction in CSVD compared with traditional statistical approaches, when using established CSVD imaging markers, along with clinical and cognitive predictors. Both survival and classification analyses were performed, and the feature importance for each model was analysed. Furthermore, the imaging data from two of the three cohorts were gathered for investigating whether DL survival models leveraging the original MRI images can improve dementia prediction in CSVD compared with using established CSVD imaging markers. Specifically, T1, FLAIR, and mean diffusivity (MD) images were utilised along with clinical and cognitive features. Both unimodal and multimodal DL survival models were trained with different strategies and compared against Cox regression. Results The major output from Part II was the Lenticulostriate artery Ultra-high-field Morphology Extraction and quantificatioN (LUMEN) pipeline. The finetuned DS6 achieved the best segmentation Dice score (>0.8), and both DS6 and nnU-Net outperformed the classical MSFDF. An easily deployable Jupyter notebook interactively linked to 3D Slicer was developed for LSA quantification. The 3D LSA morphological metrics extracted by LUMEN provide a more faithful characterisation of LSA morphology than those from 2D MIP analysis. However, no significant differences in 3D LSA morphological metrics were found across lacunar stroke patients, non-lacunar stroke controls and healthy controls. The clinical significance of extracted metrics remains to be further investigated. Part III found that when using established CSVD imaging markers along with clinical and cognitive predictors, ML methods performed similarly to Cox and logistic regression, and prediction was dominated by global cognition and age. However, when applied to the original MRI images, DL methods could improve the survival prediction of dementia compared with Cox regression using conventional CSVD imaging markers. MD was the most predictive modality among all, and multimodal DL model trained end to end from scratch achieved the best test c-index. Conclusions In conclusion, this thesis demonstrated that ML can be applied to automate the assessment of 3D LSA morphology for investigating CSVD pathophysiology, and that ML can also improve dementia prediction in CSVD when applied directly to MRI images. Overall, these findings highlight the promising potential of ML-based neuroimage analysis in improving both the assessment and prognosis of CSVD."],"dc:format.checksum.md5":["9f2d2bc51c451e5f4b9725ab074ee4ca","87eda9de84448d1f82354d60eee3eb5f"],"dc:identifier.doi":["https://doi.org/10.17863/CAM.125288"],"dc:identifier.uri":["https://www.repository.cam.ac.uk/bitstreams/4feb2db3-9fc6-4c77-908f-b71c1b24e445/download"],"dc:language":["eng"],"dc:publisher.institution":["University of Cambridge"],"dc:relation.isreferencedby.uri":["https://www.repository.cam.ac.uk/handle/1810/395952"],"dc:rights":["https://www.repository.cam.ac.uk/bitstreams/808c678f-3f26-47c3-bf89-8f1167c09a8b/download","http://purl.org/NET/rdflicense/allrightsreserved"],"dc:subject":["machine learning","MRI","cerebral small vessel disease"],"dc:title":["Applications of Machine Learning to Neuroimage Analysis in Cerebral Small Vessel Disease"],"dc:type":["Thesis"],"dc:type.qualificationlevel":["Doctoral"],"dc:type.qualificationname":["Doctor of Philosophy (PhD)"]},"updated_at":"2026-07-22T22:23:56Z"}